Biotechnology & Biotechnological Equipment

ISSN: 1310-2818 (Print) 1314-3530 (Online) Journal homepage: www.tandfonline.com/journals/tbeq20

PID Controller Tuning based on Metaheuristic

Algorithms for Bioprocess Control

Olympia Roeva & Tsonyo Slavov

To cite this article: Olympia Roeva & Tsonyo Slavov (2012) PID Controller Tuning based on
Metaheuristic Algorithms for Bioprocess Control, Biotechnology & Biotechnological Equipment,
26:5, 3267-3277, DOI: 10.5504/BBEQ.2012.0065

To link to this article: https://doi.org/10.5504/BBEQ.2012.0065

Published online: 16 Apr 2014.

Submit your article to this journal

Article views: 1089

View related articles

Citing articles: 4 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tbeq20
 Article DOI: 10.5504/BBEQ.2012.0065

BIOINFORMATICS
PID Controller Tuning based on Metaheuristic Algorithms
for Bioprocess Control
Olympia Roeva1, Tsonyo Slavov2
1
Bulgarian Academy of Sciences, Institute of Biophysics and Biomedical Engineering, Sofia, Bulgaria
2
Technical University of Sofia, Department of Systems and Control, Sofia, Bulgaria
Correspondence to: Olympia N. Roeva
E-mail: [email protected]

ABSTRACT
This paper presents an optimal tuning of a universal digital PID controller using metaheuristics as Genetic Algorithms (GA),
Simulated Annealing (SA) and Tabu Search (TS). The controllers were used to control the feed rate and to maintain the glucose
concentration at the desired set point for an E. coli MC4110 fed-batch cultivation process. The mathematical model of the
cultivation process was represented by the dynamic mass balance equations for biomass and substrate. In the control algorithm
the design measurement and process noise as well as the time delay of the glucose measurement system were taken into account.
To achieve good closed-loop system performance metaheuristics based controller tuning was done. By tuning the constants (Kp,
Ti, Td, b, c and N) in the PID controller algorithm, the controller can provide control action designed for the specific process
requirements. To evaluate the significance of the tuning procedure and controller performance different criteria were used.
Objective function values and CPU time were used as criteria to compare the performance of the three metaheuristic algorithms
– GA, SA and TS. A series of procedures for PID controller tuning were performed using competing techniques and criteria.
As a result the set of optimal PID controller settings was obtained. For a short time the controller set the control variable and
maintained it at the desired set point during the E. coli MC4110 fed-batch cultivation process. The simulation results indicate
that the proposed metaheuristic algorithms are effective and efficient, and demonstrate that the applied techniques exhibit a
significant performance improvement over classical optimization methods.

Biotechnol. & Biotechnol. Eq. 2012, 26(5), 3267-3277 The control strategy for substrate feed rate can be
Keywords: E. coli, genetic algorithm, simulated annealing, summarized in three groups: open (feedforward), closed-
loop (feedback) control and mixed (feedforward–feedback).
tabu search, PID controller, tuning, bioprocesses
A widely used controller in the feedback control of industrial
cultivation processes is the proportional-integral-derivative
Introduction (PID) controller (4, 5, 6, 9, 15, 39). The PID controller is
Fed-batch operation is often used in the industry because it often poorly tuned owing to highly changing dynamics of
allows to overcome catabolite repression or glucose effect, most bioprocesses caused by the nonlinear cell growth and
which usually occurs during production of fine chemicals the changes in the overall metabolism. Due to a change of the
such as pharmaceutical biochemicals (interleukins, insulin, system parameters, the conventional PID controllers result
interferons, enzymes and growth factors). Moreover, fed-batch in sub-optimal corrective actions and hence require retuning.
operation also gives the operator the freedom to manipulate In control design of continuous cultivation processes the
the process via the substrate feed rate. Since either nutrient controller tuning could be done with traditional methodology
overfeeding, or underfeeding, is detrimental to cell growth and (23). The models of these processes can be linearized in an
product formation, development of a suitable feeding strategy equilibrium point. Fed-batch cultivation processes, however,
control is critical in fed-batch cultivation. cannot be linearized around an equilibrium point of a system
– there is no equilibrium point. If a linear approximation is
There are commercially available controllers only for well-
found, the resulting model will be valid only for a small region
established measurement systems (such as temperature, pH and
around the linearization point. Therefore, it is necessary to use
dissolved oxygen); for substrate feed rate control there is a lack non-classical tuning methods to achieve the best overall PID
of control systems. The main reason for this are the difficulties control for the entire operating envelope of the given system.
in on-line measuring of the substrate concentration in a fast Optimization methods could be applied for the controller tuning,
and reliable way during the cultivation process. Moreover, in although the procedure is a big challenge for the conventional
principle, the substrate concentration measurement systems optimization methods. As an alternative, metaheuristics could
are characterized with significant time delay which sets the be applied for the quality controller tuning (12, 17, 18, 19, 22,
challenge to control the processes. 30).

Biotechnol. & Biotechnol. Eq. 26/2012/5 3267

 Metaheuristics obtain suboptimal solution in ordinary introduced (38). The overall results show that SA yields better
situations and optimal solution in particular. Since the performance when compared to GA. Haber et al. (13) show a
considered problem is known to be NP-complete, using strategy based on SA for the optimal tuning of a PID controller
heuristic techniques can solve this problem more efficiently. to deal with time-varying delay. The proposed strategy is
The three most well-known heuristics are the iterative compared with other classic tuning rules (the Ziegler–Nichols
improvement algorithms, the probabilistic optimization and Cohen–Coon tuning formulas). The results demonstrate
algorithms, and the constructive heuristics. In the probabilistic that SA provides an optimal tuning of the PID controller. A
optimization group, genetic algorithm (GA) based methods similar investigation is presented in (1, 3, 22). There is a lack
and simulated annealing (SA) are the most studied ones (10, of results about using SA for design of bioprocess control
13, 17, 18, 19, 20, 23, 29, 37). A relatively new metaheuristic systems.
technique is Tabu Search (TS), which has also been proven Uaraboga and Ualinli (16) describe a new method for
effective for many optimization problems (10). tuning the control parameters of a PID controller based on
The GAs are highly relevant for industrial applications the TS algorithm. Optimal PID controller parameters of an
because they are capable of handling problems with nonlinear automatic voltage regulator system, using adaptive TS (ATS)
constraints, multiple objectives, and dynamic components – algorithm, are presented (26, 27, 28). By comparison with the
properties that frequently appear in the real-life problems (11, conventional technique, the effectiveness of the anticipated
23). Since their introduction and subsequent popularization, scheme is confirmed. ATS algorithm for the PID controller
the GAs have been frequently used as alternative optimization design problems is applied by Puangdownreong et al. (31), too.
tools to the conventional methods (11, 29), have been Directed TS algorithm is used to optimize the gains of a PID
successfully applied in a variety of areas, and still find controller for a class of time delay systems (8). No publications
increasing acceptance. about applications of TS for bioprocess control systems design
Another competing metaheuristic algorithm is SA (20). SA could be found.
is a stochastic relaxation technique that uses the Metropolis This work focuses on designed glucose concentration
algorithm based on the Boltzmann distribution in statistical control system based on metaheuristic techniques tuned
mechanics, for solving nonconvex optimization problems. SA universal digital PID controller. E. coli MC4110 fed-batch
can deal with arbitrary systems and cost functions; statistically cultivation process is considered.
guarantees finding an optimal solution (SA has the ability to
avoid getting stuck at local minima); guarantees a convergence Materials and Methods
upon running sufficiently large (infinite) number of iterations; Process model
is relatively easy to code, even for complex problems. This The mathematical model (Eq. 1–Eq. 5) is based on a real fed-
makes annealing an attractive option for optimization problems. batch cultivation process of E. coli MC4110. The cultivation
TS is a metaheuristic method that guides a local heuristic process was carried out in the Institute of Technical Chemistry,
search procedure to explore the solution space beyond local Hannover University, Germany. The cultivation conditions
optimality (10). It is also different from the SA and GA because and data measurements are previously discussed (2, 36). The
the TS includes a memory mechanism. Most TS versions can model is presented as:
be characterized by the following two important properties:
complementing local search (the neighborhood concept), and
prohibiting moves that have been previously selected (the (Eq. 1)
adaptive memory concept). TS is very general and conceptually
t
much simpler than either SA or GA. x ( t ) =  X ( t ) S ( t ) V ( t ) m max ( t ) 
(Eq. 2)
Controller tuning procedures using metaheuristics have
been presented in a few articles. Considering controllers for f ( x, F ) =
bioprocess control, to the best of our knowledge, there is a lack  S (t ) F (t ) 
of such publications.  mmax ( t ) X (t ) − X (t ) 
 k S + S ()
t V ()
t 
GA has been applied for tuning of a PID controller for a  
continuous model bioreactor (23). It is also shown that applying 1 S ()t F ()t
= −
 YS / X
mmax ( t )
kS + S ( t )
X (t ) +
V (t )
( Sin − S ( t ) ) 

GA global and local optimal solution can be simultaneously
 
achieved (17, 18, 19). The generality of the GA, combined  F (t ) 
with its intuitiveness, fast convergence, modest processing  0 
 , (Eq. 3)
requirements and, most importantly, minimal system specific
information result in increased use of this technique for tuning h = [ 0 1 0 0]
of PID controllers (12, 21, 24, 37). , (Eq. 4)
The application of SA to the problem of tuning the PI t
ç ( t ) = η X ( t ) η S ( t ) 0 η mmax ( t ) 
controller for a linearized coupled tank liquid level control is , (Eq. 5)
3268 Biotechnol. & Biotechnol. Eq. 26/2012/5
where X is the concentration of biomass, g·L-1; S is the substrate Select P(i) from P(i – 1) (select a subpopulation for
(glucose) concentration, g·L-1; F is the feed rate, L·h-1; V is the offspring reproduction)
Recombine P(i) (recombine the genes of selected parents)
bioreactor volume, L; Sin is the substrate concentration of the Mutate P(i) (perturb the mated population stochastically)
feeding solution, g·L-1; mmax is the maximum growth rate, h-1; kS Evaluate P(i) (evaluate its new fitness)
is the saturation constant, g·L-1; YS/X is the yield coefficient, -; end
ηX – biomass concentration process noises, g·L-1; ηS – substrate end
concentration process noises, g·L-1; η mmax – maximum growth Simulated annealing
rate process noises, h-1; ξ (t ) is measurement noise, g·L-1.
For SA described in greater details see (39). The trial point
Based on real experimental data (feeding rate data and off- distance distribution could be set as function that generates a
line measurements of biomass and on-line data of substrate point based on the current point and the current temperature
[glucose] measurements) and GA identification procedure using different distributions. Here the Boltzmann distribution
the following numerical values of the model parameters were was used:
obtained (36): kS = 0.012 g·L-1, YS/X = 0.5.
The initial process conditions were (2): Pr { E} = exp { E ( i ) / kT } / Z (T )
, (Eq. 8)
t0 = 6.68 h, X(t0) = 1.25 g·L , S(t0) = 0.8 g·L , Sin = 100 g·L .
-1 -1 -1
where E is the system energy; k is Boltzmann’s constant
The model inaccuracy is modeled via zero mean white (k = 1.380650×10-23 m2·kg·s-2·k-1); T is the system temperature,
Gausian noises. The corresponding variances were (2): K; Z(T) is a normalization function:
Dηγ = 0.001 g2·L-2·h-1, Dηγ = 0.001 g2·L-2·h-1, Dη = 0 L2·h-
X Q
Z (T ) = ∑ exp { E ( h ) / kT }
S
3
and Dηmmax = 0.05 L·h . -3
h . (Eq. 9)
The specific grow rate ì ( t ) is described by Monod The initial temperature was 100 K.
kinetics: The probability of acceptance is (25):

S (t ) 1
m ( t ) = mmax ( t ) 1 + exp ( D / max (T ) )
kS + S ( t ) . (Eq. 10)
. (Eq. 6)
The function that the algorithm uses to update the
Genetic algorithm temperature is:
GAs are described in greater details by Chipperfield et al. (7).
In the GA applied here the individual fitness, F(xi), of each T = T0 0.95r , (Eq. 11)
individual is computed as the individual’s raw performance,
f(xi), relative to the whole population, i.e., where r is the annealing parameter:
Nind  max ( s j ) 
T j
F ( xi ) = f ( xi ) / ∑ f ( xi ) ri = log  0 
 Ti si 
i =1 , (Eq. 7)  , (Eq. 12)
where f is the objective function, F is the resulting relative
where ri is the annealing parameter for component i, T0 is the
fitness, Nind is the population size and xi is the phenotypic value
initial temperature of component i, Ti is the current temperature
of individual i.
of the component i, si is the gradient of the objective in direction
To solve the considered optimization problem the population i times difference of bounds in direction i. In the considered
size was chosen to be 100 after several algorithm performance algorithm the number of points accepted before reannealing
pre-tests. In the same manner the number of generations was 100.
was set at 100. Here double point crossover with crossover
The algorithm stops when the average change in the
probability of 0.7 was assumed. Mutation is randomly applied
objective function is sufficiently small with respect to the
with low probability, typically in the range 0.001 and 0.01
predefined tolerance.
(used here). A generation gap of 0.97 was chosen and fitness-
based reinsertion was used. The SA algorithm can be described by the following
schema:
The following is a pseudocode of a GA: Find initial solution (by generating it randomly)
begin Set initial value for the control parameter T = T0
i = 0 (set generation number to zero) Set a value for r, the rate of cooling parameter
Initial population P(0) (initialize a usually random population j=0
of individuals) Generate (at random) a new solution S’
Evaluate P(0) (evaluate fitness of all initial individuals of Calculate the difference in cost:
population) D = cost(S’) – cost(S)
while (not done) do (test for termination criterion [time, fitness, Examine the new solution and decide:
etc.]) accept or reject
begin If accepted, it becomes the current solution; otherwise, keep the
i = i + 1 (increase the generation number)
Biotechnol. & Biotechnol. Eq. 26/2012/5 3269
 old one; j = j+1 the accuracy of the decisions and the computing time could be
Reduce the temperature and generate a new solution optimized to a not inconsiderable degree. The parameters of the
Until some stopping criterion applies
regarded algorithms were tuned based on several pre-tests for the
Tabu search PID controller tuning problem considered here. Following the
TS has similar characteristics to SA, i.e. starting from one tuning procedures the main metaheuristic algorithm parameters
initial solution and generating a new solution to search through were set to the optimal settings, presented in the next four tables.
its neighborhood. A detailed description of TS algorithm is The GA operators and parameters are summarized in
given by Gendreau and Potvin (12). Table 1 and Table 2. These are the most appropriate GA
The following is a pseudocode of TS: parameters and operators for the problem considered here
Step 1. Initialization based on our earlier research (32, 33, 34, 35) too. The SA
Set k = 1 algorithm parameters are summarized in Table 3 and the TS
Generate initial solution S0
Set S1 = S0, then G(S1) = G(S0)
algorithm settings, in Table 4.
Step 2. Moving
Select Sc from neighborhood of Sk
Table 1
IF move from Sk to Sc is already in TL THEN Genetic algorithm operators
Sk+1 = Sk
GOTO Step 3 Operator Type
END IF
IF G(Sc) = G(S0) THEN Encoding Binary
S0 = Sc Crossover Double point
END IF Mutation Bit inversion
Delete the TL move in the bottom of TL
Add new Tabu Move in the top of TL Selection Roulette wheel selection
GOTO Step 3 Fitness function Linear ranking
Step 3. Next iteration
Set k = k + 1 Table 2
IF k = N THEN Genetic algorithm parameters
STOP
ELSE
GOTO Step 2 Parameter Value
END IF Generation gap 0.97
Computer specifications Crossover rate 0.70
The computer specifications to run all optimization procedures Mutation rate 0.05
were Intel® Core™2 Quad CPU Q8200 2.33 GHz, 4 GB Memory, Precision of binary representation 20
Windows XP operating system and Matlab 7.5 environment. Number of individuals 100
Number of generations 100
Results and Discussion Table 3
In order to prevent E. coli growth inhibition due to substrate Simulated annealing parameters
excess, the glucose concentration has to be controlled at low
levels. Two control systems are considered. In the first control Parameter Value
system (simpler case), measurement and process noise are taken Reannealing interval 100
into account. In the second control system (more complex case),
Initial temperature 100
a Smith Predictor (SP) structure is used. Moreover, to reduce
Maximum number
the influence of glucose measurement system time delay, a
of objective function 3000*number of variables
correction in measured glucose is proposed. To achieve good
evaluations allowed
closed-loop system performance GA, SA and TS based optimal
Termination tolerance on
controller tuning procedures are applied. Objective function 10-6
function value (TolFun)
values and CPU times were used as criteria to compare the
Step length equals uniformly
performance of the three metaheuristic algorithms. A series of
Annealing function random current temperature
tuning procedures for PID controllers tuning, using competing
direction
techniques and different criteria, are performed. As a result the
set of optimal PID controller’s settings are obtained. Acceptance function See Eq. 10
Number of iterations over
Metaheuristic algorithms settings which average change in fitness
500*number of variables
In order to increase the performance of the three competing function value at current point
algorithms (GA, SA and TS) it is necessary to provide is less than TolFun
adjustments of their parameters depending on the considered Function used to update
See Eq. 11
problem. With the appropriate choice of algorithm settings temperature schedule

3270 Biotechnol. & Biotechnol. Eq. 26/2012/5

 Table 4 worst results of the 30 runs, for the objective function value I
Tabu Search parameters and execution time (CPU time) were observed.
Two glucose concentration control systems of an E. coli
Parameter Value fed-batch cultivation process were considered.
0.1*min(upper limits-lower Case 1: In this case a universal digital PID controller is
Simplex edge length (SEL)
limits) used. A typical structure of a PID control system is shown in
Ratio of accepting Fig. 1. The error signal e(t) is used to generate the P, I, and
1
diversification point D modes, with the resulting signals weighted and summed to
Maximum number of iteration 100*number of variables form the control signal u(t) applied to the plant model (Eq. 1–
Termination accuracy 10-6 Eq. 5).
Number of tabus
5*number of variables
in TL (TN)
TR radius 0.01*SEL
Region radius 2*SEL
Number of best points saved
TN/2.5
in TL
To evaluate the significance of the tuning procedure and Fig. 1. Structure of the designed control system.
controller performance four criteria were used: integrated
square error (IISE); integrated absolute error (IIAE); integrated Introducing coefficients b, c and first-order low pass
time-weighted absolute error (IITAE) and integrated squared filter in D mode leads to negligibly more complex controller
time-weighted error (IISTE): but sufficiently improves the control system performance.
The coefficient b, (0 ≤ b), is used to weight out the r(t) in P
T T
mode of the controller and the coefficient c (c ≤ 1) is used
I ISE = ∫ e ( t ) dt I IAE = ∫ e ( t ) dt
2

to weight out the r(t) in D mode of the controller. Typically

0 , 0 ,
in industrial applications b and c are chosen to be equal to 0
T T or 1. Using a first-order low-pass filter reduces the influence
I ITAE = ∫ t e ( t ) dt I ISTE = ∫ t 2 e ( t ) dt
2 2
of measurement noise. Considering the real applications, a
0 , 0 , digital PID controller is usually implemented. There are many
where the error e is the difference between the set-point and techniques for discretization. Here for discretization of the
the estimated substrate concentration (SSP – S), t is time, T is PID controller the backward Euler method (14) was used.
the end time of the cultivation. The mathematical description of discrete-time universal
For the E. coli MC4110 cultivation process considered here PID controller is:
the desired set-point is at SSP = 0.1 g·l-1 glucose concentration u (k ) = u p (k ) + ui (k ) + ud (k )
(2). Concentrations above this value lead to a substrate , (Eq. 13)
inhibition of the process, i.e. to a negative effect on the
u p (k ) = K p [br (k ) − y (k ) ]
productivity and yield of a desired cultivation product. , (Eq. 14)
Numerical results and discussion ui (k ) = ui (k − 1) + bi1 [ r (k ) − y (k ) ] +
A series of tuning procedures for PID controller tuning, + bi 2 [ r (k − 1) − y (k − 1) ]
using competing techniques and different criteria, were , (Eq. 15)
performed. Because of the stochastic characteristics of the ud (k ) = ad ud (k − 1) +
applied algorithms a series of 30 runs for each algorithm were
+ bd [ cr (k ) − cr (k − 1) − y (k ) + y (k − 1) ]
performed. To compare the GA, SA and TS, the mean, best and , (Eq. 16)

Table 5
Controller parameters tuned with GA, SA and TS algorithms

GA SA TS
Parameter
Mean Best Worst Mean Best Worst Mean Best Worst
Kp 0.4002 0.4000 0.4000 0.4009 0.4000 0.4000 0.4925 0.5252 0.7904
Ti 0.9999 1.0000 1.0000 0.9346 0.9930 1.0000 0.7182 0.6281 0.9779
Td 0.0030 0.0030 0.0030 0.0030 0.0030 0.0030 0.0392 0.0239 0.0302
IISE 16.1945 16.1943 16.1943 16.1944 16.1943 16.1949 16.3431 16.1878 16.6157
CPU time, s 267.5859 200.2656 353.3438 192.5766 176.4219 309.6563 43.3625 39.2813 47.7031
Biotechnol. & Biotechnol. Eq. 26/2012/5 3271
where bi1 = K pT0 / Ti , bi2, ad = Td / (Td + NT0 ) , concentration) is displayed. It is seen that for a short time the
controllers set the control variable and kept it at the desired set
bd = K pTd N / (Td + NT0 ) , r(k) is a reference signal, y(k) is
point of 0.1 g·L-1 during the process.
output signal, u(k) is control signal, Kp is proportional gain, Ti
is integral time, Td is derivative time, Td/N is time constant of Control variable
first-order low-pass filter, T0 is sample time. 1

The optimal value of the PID controller parameters (Kp, 0.2

Genetic algorithm, Simulated Annealing

Glucose concentration , [g/l]

Ti, Td) was found using GA, SA and TS. By their tuning the 0.8 0.15 Tabu search

controller can provide control action designed for specific

process requirements. The range of the tuning parameters was 0.6
0.1

considered as follows: Kp ∈ [0, 2], Ti ∈ [0, 1], Td ∈ [0, 0.1]. 0.05

After several runs the range for the parameters was specified
0.4
to: Kp ∈ [0.4, 2], Ti ∈ [0.005, 1] and Td ∈ [0.003, 0.1]. The 0
9 9.5 10 10.5 11 11.5 12
coefficients b, c and N were set to the following values: b = 1,
c = 1 and N = 30. As a result of the metaheuristics tuning the 0.2
optimal PID controller settings were obtained. The numerical
values of the controller parameters for the four criteria were 0
quite similar for each algorithm. The considered objective 6 8 10 12 14 16
Cultivation time [h]
functions reflect the performance of the PID controller in a
similar way. Due to this fact Table 5 presents only the results Fig. 2. Glucose concentration during the process (Case 1).
for the case of IISE. For better visualization and analysis in Fig. 2 the substrate
The results from GA and SA tuning showed that the concentrations between 9 h and 12 h are presented. In Table 6
algorithms produce the same estimations with more than 99 % the mean value of S (mS), the standard deviation of S (σS) and
coincidence. In the case of TS there was a big difference between the maximum deviation of S from the set point (σmaxS) are
the obtained controller parameters values. The obtained IISE for shown. The best results were observed for the TS algorithm,
all algorithms are representative and sophisticated controller although the differences are negligible.
performance indices. At the same time they are almost equal: Fig. 3 and Fig. 4 show the biomass concentration and feed
16.1943 for GA, 16.1943 for SA and 16.1878 for TS. It can be rate profile during the E. coli cultivation process. Optimal
concluded that the results are functionally close. controller parameters led to a high biomass concentration
in the end of the cultivation process (35 g·L-1). For the three
Table 6
algorithms the feed rates did not reach the actuator limitations.
Some statistical characteristic of S (Case 1)
It can be seen that for TS the feed rate profile had smaller
oscillations than those for GA and SA. This is the most
mS σS σmaxS
practically applicable case for the feed rate pump.
GA 0.1031 0.0325 0.064
SA 0.1031 0.0325 0.064 Biomass concentartion
TS 35
0.1020 0.0297 0.042 29

The same results for the controller parameters were 30 28.9

Biomass concentartion , [g/l]

obtained by applying GA and SA. It was clearly shown that 28.8

the execution time of the GA and SA are similar too. The best 25 Tabu Search
28.7
result from SA was achieved for 176.4219 s compared to GA, 20 28.6
Simulated Annealing,
Genetic Algorithm
200.2656 s. A larger difference was observed in the average
28.5
results: 192.5766 s for SA and 267.5859 s for GA. In contrast the 15 14.9 14.92 14.94 14.96

TS obtained results for about four times less computation time

in comparison with the best results given by SA, respectively 10
39.2813 s and 176.4219 s. This is due to the different nature of 5
the competing algorithms. For SA the execution time depends
only on the problem size and the number of iterations. For GA 0
6 8 10 12 14 16
the execution time is does not merely depend on the problem Cultivation time [h]
size and the number of iterations, but also depends on the
Fig. 3. Biomass concentration during the process (Case 1).
population size and the representation. TS is much simpler
than either SA or GA but at the same time, more powerful For PID controller parameter tuning of the considered
using the adaptive memory concept. control system the three metaheuristic algorithms (GA, SA
The following figures show some results for the control and TS) showed identical performance. For comparison of the
system performance. In Fig. 2 the control variable (glucose applicability of these algorithms in more realistic conditions

3272 Biotechnol. & Biotechnol. Eq. 26/2012/5

of E. coli fed-batch cultivation process, another more complex In the designed control system the SP structure was used
control system was considered. for reducing the influence of the measurement system delay.
For the estimation of the process variables which are used
Feed rate
1 to form the feedforward term of control signal, an extended
0.25
Simulated Annealing, Kalman filter (EKF) was designed. To form the feedback term
0.2
Genetic Algorithm
of control signal the same PID controller as that in Case 1 was
0.8
used. The structure of the control system is shown in Fig. 5.
0.15
The block “Nonlinear process model” predicts the non-
Feed rate , [l/h]

0.6 delayed model output by:

0.1 Tabu Search

x& m ( t ) = f m ( x m , F )
11 11.2 11.4 11.6 11.8 12

0.4
Sm ( t ) = h m xm ( t )
0.2 mm ( t ) X m ( t )
SCORm ( t ) = S m ( t ) + Dt ,
YS / X
0 Sm ( t )
6 8 10 12 14 16 mm ( t ) = mmax
Cultivation time [h]
m
kS + Sm ( t )
(Eq. 18)
Fig. 4. Feed rate profiles during the process (Case 1).
t
Case 2: In this case the process model (Eq. 1–Eq. 5) is x m ( t ) =  X m ( t ) S m ( t ) Vm ( t ) 
, (Eq. 19)
extended with equation for the correction of measured glucose
concentration: fm ( xm , F ) =

ì (t ) X (t )  F (t ) 
SCOR ( t ) = S ( t ) + Dt  mm ( t ) X m ( t ) − X m (t ) 
YS / X  V m ( )
t 
. (Eq. 17)  
1 F ( t )
= −
 YS / X
mm ( t ) X m ( t ) +
Vm ( t )
( Sin − S m ( t ) ) 

η ξ
Freal  
F (t )
r e e* u fbreal Cultivation S
PID
Controller process  
u ff real
Extended
 
Kalman Filter  , (Eq. 20)
Feed forward Ŝ
control
Sm
where Xm is the concentration of biomass, evaluated by the
SCORm
Nonlinear Xm em model, g·L-1; Sm is the delayed concentration of substrate
process model
mmax m
evaluated by the model, g·L-1; Vm is the bioreactor volume
Vm
evaluated by model, L; m max m is the model maximum growth
rate, h-1, SCORm is the nondelayed concentration of substrate
predicted by the model, g·L-1. Here m max m = 0.5 h-1.
Fig. 5. Structure of the control system. The feedback control algorithm (PID controller) is
described as follows:

Table 7
Controller parameters tuned with GA, SA and TS algorithms (Case 2)

GA SA TS
Parameter
Mean Best Worst Mean Best Worst Mean Best Worst
Kp 0.0190 0.0200 0.2937 0.3452 0.8592 6.58*10-5 0.4320 0.5073 0.0106
Ti 0.0294 0.0368 0.0259 4.5981 6.7438 6.58*10-5 1.0071 1.3819 0.3375
Td 0.0568 0.0558 0.0053 0.0235 0.0069 11.9000 0.0017 0.0009 0.1381
b 1.8696 0.7980 2.1480 2.1761 2.6000 1.7500 0.8381 0.6494 3.2767
c 0.9464 0.9998 0.3737 0.5428 0.3488 1.2000 1.4392 1.6851 0.6392
N 7.8241 9.9300 6.1798 21.9871 19.7200 28.2900 6.3419 7.5892 2.5618
IISE 16.6972 16.6800 16.7087 18.2213 16.8131 20.3914 17.3078 17.166 17.6968
CPU time, s 3173.4 2535.5 3985.2 1502.4 1783.6 1365.6 995.7 1008.5 987.2

Biotechnol. & Biotechnol. Eq. 26/2012/5 3273

 Kp Freal ( k ) = u fbreal ( k ) + u ff ( k )
u fb ( s ) = K p ( be ( s ) − SCORm ( s ) ) + e* ( s ) + , (Eq. 31)
Ti
where
Td s
+
Ts
( ce ( s ) − SCORm ( s ) ) , u fbreal ( k ) = u preal ( k ) + uireal ( k ) + udreal ( k )
1+ d . (Eq. 32)
N (Eq. 21)
To provide control action designed for specific process
where ufb(s) is the feedback term of control variable, L·h-1. requirements, it is required to tune the PID controller
The control error e*(s) is the difference between the error parameters. The controller parameters are: Kp, Ti, Td, b, c and
e(s) (difference between set-point [reference signal] r(s) and N. The range of the tuning parameters is considered as follows:
the model error em(s)) and the corrected glucose concentration Kp, Ti, Td ∈ [0, 100], b, c ∈ [0, 3], N ∈ [5, 30].
SCORm(s) (Eq. 18); the model error em(s) is the difference For comparison of the GA, SA and TS, the mean, best and
between the measured glucose concentration S(s) and the one worst results of the 30 runs, for the J value and execution time
Sm(s) evaluated by the process model. were again observed. As in Case 1 only the results for the case
e* ( s ) = e ( s ) − SCORm ( s ) , of IISE are presented (Table 7).
e ( s ) = r ( s ) − em ( s ) , In this case the results showed that the algorithms produce
quite different controller parameters. This was more marked
em ( s ) = S ( s ) − S m ( s ) .
(Eq. 22) in the case of SA. The resulting best values of the objective
The mathematical description of the designed digital PID function were similar for GA and SA, respectively 16.6800
controller is: and 16.8131. The TS gave a slightly worse result – 17.166.
At the same time the resulting control systems performances
u fb ( k ) = u p ( k ) + ui ( k ) + ud ( k ) were quite different. Until about 12 h of cultivation the three
, (Eq. 23)
controllers kept the control variable at the desired set point
u p ( k ) = K p ( be ( k ) − SCORm ( k ) ) in a similar manner. The control variable is given in Fig. 6.
, (Eq. 24) As it can be seen form Fig. 6B, TS set the control variable
with 20 % overshot; whereas SA, with 13 %; and GA, with
ui ( k ) = ui ( k − 1) + bi1 ( e ( k ) − SCORm ( k ) ) +
less than 10 %. Until 12 h the three controllers maintained
+ bi 2 ( e ( k − 1) − SCORm ( k − 1) ) stable glucose concentration. After 12 h the controllers tuned
, (Eq. 25) by SA and TS were unable to keep the glucose concentration
ud ( k ) = ad ud ( k − 1) + bd (ce ( k ) − ce ( k − 1) − at 0.1 g·L-1. Only the GA controller successfully kept the set
point until the end of the cultivation process (Fig. 6C). The
− SCORm ( k ) + SCORm ( k − 1))
. (Eq. 26) resulting oscillations in the case of SA and TS controllers are
The control variable used to control the feed rate is: inacceptable and will lead to cultivation process break.
In Table 8 the mS, σS and σmaxS values are shown. The mean
1 Vm ( k ) m m ( k ) X m ( k ) value results indicate that the GA controller maintained the
F ( k ) = u fb ( k ) +
YS / X Sin − SCORm glucose concentration with highest accuracy. The maximum
. (Eq. 27)
deviation of the GA controller was about 20–30 times smaller
The error em(k) is calculated as: than those of the other two controllers.
em (k ) = Sˆ ( k ) − Sm ( k ) Table 8
. (Eq. 28)
Some statistical characteristic of S (Case 2)
Based on discretization of the process model (Eq. 1–Eq. 5)
the following EKF was obtained:
mS σS σmaxS
(
xˆ ( k + 1) = fd ( xˆ ( k ) ) + K eKf ( k + 1) S ( k + 1) − hfd ( xˆ ( k ) ) , ) GA 0.0984 0.0065 0.0100
SA 0.1214 0.0266 0.2105
Sˆ ( k + 1) = hxˆ ( k + 1) , TS 0.0978 0.0497 0.3312
(Eq. 29)
Fig. 7 and Fig. 8 show the biomass concentration and feed
xˆ (0) = [1.25 0.8 1.35 0.5] ,
t
rate profile during the process. The control system tuned by
GA yielded a high biomass concentration at the end of the
fd ( xˆ ( k ) ) = xˆ ( k ) + T0 f ( xˆ ( k ) ) E. coli cultivation process (48.4 g·L-1). The results in Fig. 7
. (Eq. 30)
show that the TS controller achieved 47.6 g·L-1, but in real
where xˆ ( ⋅) and Ŝ ( ⋅) are the estimates of x(⋅) and S (⋅) ; conditions such a concentration will not be obtained due to
K eKf ( ⋅) is the EKF gain. necessary cultivation process break somewhere around 14 h. If
this is taken into account, the final biomass concentration will
Finally, the real control variable is:
be 24 g·L-1. The situation is the same for the SA controller. The

3274 Biotechnol. & Biotechnol. Eq. 26/2012/5

 Control variable Biomass concentration
1 50
Glucose concentration (corrected), [g/l]

Genetic algorithm
0.8

Biomass concentration, [g/l]

Tabu search
Tabu search 40
0.6 Simulated annealing

30
0.4

0.2
20

0 Simulated annealing

Genetic algorithm 10
-0.2

-0.4 0
6 8 10 12 14 16 6 8 10 12 14 16
A Cultivation time [h] Cultivation time [h]
Fig. 7. Biomass concentration during the process (Case 2).

Control variable Control signal

0.8
Glucose concentration (corrected), [g/l]

0.14 0.7

0.13 0.6
Genetic algorithm Simulated annealing
Feed rate, [l/h]

0.12 0.5
Simulated annealing

0.11 0.4

0.3
0.1
Tabu search
0.2
0.09
Tabu search 0.1
0.08
Genetic algorithm
0
7.5 8 8.5 9 9.5 6 8 10 12 14 16
B Cultivation time [h] Cultivation time [h]
Fig. 8. Feed rate profiles during the process (Case 2).

Control variable Specific growth rate

0.4 1.4
Glucose concentration (corrected), [g/l]

0.35
Tabu search 1.2
Specific growth rate, [1/h]

0.3
1
Simulated annealing
0.25 Tabu search
Simulated annealing 0.8
0.2
0.6
0.15

0.1 0.4

0.05 0.2
Genetic algorithm Genetic algorithm
0 0
12 13 14 15 16 6 8 10 12 14 16
C Cultivation time [h] Cultivation time [h]
Fig. 6. Glucose concentration during the process (Case 2). (A) (B) (C) Fig. 9. Resulting specific growth rates during the process (Case 2).

Biotechnol. & Biotechnol. Eq. 26/2012/5 3275

resulting oscillations after 12 h will break the process and the In the second case a universal digital PID controller with
final biomass concentration will be 12 g·L-1. It is clear that only the SP structure was used. As well as measurement and process
the PID controller tuned by GA could be used for prolonged noise, the time delay of glucose measurement system was
control of glucose concentration during the considered taken into account too. Here the competing algorithms showed
cultivation process. considerably different control system performance. In this
The resulting specific growth rates are presented in Fig. 9. case, which is more realistic, the PID controller tuned by GA
The unsuccessful performance of SA and TS controllers lead could be used for prolonged control of glucose concentration
to unrealistic dynamics of the specific growth rates after 12 h during the E. coli fed-batch cultivation process in comparison
and 14 h of cultivation time, respectively. with the other two controllers. As a result, a 2 to 4 times
higher final biomass concentration was achieved using the GA
Conclusions controller instead of SA and TS controllers. The SA and TS
This paper focuses on an optimal tuning of PID controllers controllers failed to maintain the control variable in the final
using metaheuristics such as GA, SA and TS. The controller 3–5 h of the cultivation process.
was used to control the feed rate and maintain the glucose In summary, the obtained results show that the PID
concentration at the desired set point for an E. coli MC4110 controller tuning using metaheuristics (GA, SA and TS) can
fed-batch cultivation process. The mathematical model of the be considered an effective methodology for achievement of
cultivation process is represented by the dynamic mass balance high quality and better performance of the designed control
equations for biomass (bacteria) and substrate (glucose). system for cultivation processes.
In order to prevent the growth inhibition based on substrate
excess, the glucose concentration has to be controlled at low Acknowledgements
levels. For the cultivation process considered here the desired This work was partially supported by the National Science
set-point was at SSP = 0.1 g·L-1 glucose concentration. Fund of Bulgaria, Grants DMU 02/4 “High Quality Control
To achieve good closed-loop system performance of Biotechnological Processes with Application of Modified
metaheuristics based controller parameters (Kp, Ti, Td, b, c and Conventional and Metaheuristics Methods” and DID 02-
N) tuning was done. To evaluate the significance of the tuning 29 “Modeling Processes with Fixed Development Rules
procedure and controller performance, four criteria were used: (ModProFix)” and by the European Social Fund under
integrated square error (IISE), integrated absolute error (IIAE), Operative Programme “Human Resources Development”,
integrated time-weighted absolute error (IITAE), and integrated Grant BG051PO001-3.3.05-001 (DO2-522).
squared time-weighted error (IISTE). The objective function
values and CPU time were used as criteria to compare the
performance of the three metaheuristic algorithms (GA, SA REFERENCES
and TS). 1. Abido M.A. (2000) Electrical Power and Energy Systems,
A series of tuning procedures for PID controller tuning, 22, 247-258.
using competing techniques and different criteria, were 2. Arndt M., Hitzmann B. (2001) In: 8th IFAC Int. Conf. on
performed. Two different glucose control systems were Comp. Appl. in Biotechn., Quebec City, Canada, 425-429.
considered. In the first case a universal digital PID controller 3. Asharf S., Kumar S. (2008) International Journal of
was used. In control system algorithm design measurement Electrical, Computer, and Systems Engineering, 1(2), 71-
and process noise were taken into account. As a result a set 80.
of optimal GA, SA and TS PID controller parameters were
4. Astrom K., Hagglund T. (2006) Advanced PID Control,
obtained. For a short time the controllers set the control
ISA – The Instrumentation, System and Automation
variable and maintained it at the desired set point during the
Society, NC, USA.
cultivation process. The compared metaheuristic approaches
and statistical results show that the considered searchers 5. Astrom K., Hagglund T. (1995) PID Controllers: Theory,
found comparable solutions for the given criteria. Although Design and Tuning, Instrument Society of America,
the criteria values were similar, the resulting dynamics of the Research Triangle Park, NC, USA.
control variable obtained by TS were smoother in comparison 6. Astrom K., Hagglund T., Hang C.C., Ho W.K. (1993)
with those of the the GA and SA. The obtained values for the Control Eng. Pract., 1, 699-714.
standard deviations of the control variable were sufficiently 7. Chipperfield A., Fleming P., Pohlheim H., Fonseca C.
small. The solution times used by the three heuristics were also (1994) Genetic Algorithm Toolbox User’s Guide, Version
compared. The considered metaheuristics obtained solutions 1.2, Department of Automatic Control and Systems
use different CPU times. The applied TS procedure was 4 to 5 Engineering, University of Sheffield, UK.
times faster than the SA and GA ones. Generally, in this case 8. Ganapathy K., Jovitha J. (2010) International Journal of
the three metaheuristic algorithms (GA, SA and TS) showed Soft Computing, 5(4), 177-184.
identical performance.

3276 Biotechnol. & Biotechnol. Eq. 26/2012/5

9. Garipov E. (2006) PID Controllers, Automatics and 25. Mathworks. How Stimulated Annealing Works (Accessed:
Informatics, Vol. 3, Sofia. (In Bulgarian) Sept 2012) <http://www.mathworks.com/help/gads/how-
10. Glover F., Laguna M. (1997) Tabu Search, Kluwer simulated-annealing-works.html>
Academic Publishers, Boston. 26. Oonsivilai A., Marungsri B. (2007) In: 7th WSEAS
11. Goldberg D.E. (1989) Genetic Algorithms in Search, International Conference on Power Systems, Beijing,
Optimization and Machine Learning, Addison Wesley China, 42-47.
Longman, London. 27. Oonsivilai A., Padej P.-L.-O. (2008) WSEAS Transactions
12. Gendreau M., Potvin J.-Y. (2005) In: Search on Power Systems, 3(6), 495-506.
Methodologies, Introductory Tutorials in Optimization and 28. Oonsivilai A., Padej P.-L.-O. (2008) In: 12th WSEAS
Decision Support Techniques (E. K. Burke, G. Kendall, International Conference on Computers, 987-992.
Eds.), Springer, USA, 165-186. 29. Parker B.S. (1992) Demonstration of using Genetic
13. Haber R.E., Haber-Haber R., del Toro R.M., Alique Algorithm Learning, Information Systems Teaching
J.R. (2007) LNCS, 4507, Springer, Heidelberg, 1155-1162. Laboratory.
14. Heath M.T. (2002) Scientific Computing: An Introductory 30. Paz-Ramos M.A., Torres-Jimenez J., Quintero-
Survey, (Second Edition), McGraw-Hill, New York. Marmol-Marquez E., Estrada-Esquivel H. (2004)
15. Johnson M.A., Moradi M.H. (2005) PID Controller LNCS, 3103, Springer, Heidelberg, 1-10.
Design, Springer-Verlag, London, pg. .. 31. Puangdownreong D., Kulworawanichpong T., Sujitjorn
16. Karaboga D., Kalinli A. (1996) In: IEEE Int. Conf. on S. (2004) In: IEEE TENCON, 4, 451-454.
Systems, Man, and Cybernetics, 1, 134-136. 32. Roeva O. (2006) Biotechnol. Biotech. Eq., 20(1), 202-209.
17. Kim D.H. (2003) In: Computational Intelligence for 33. Roeva O. (2008) Advanced Topics on Evolutionary
Measurement Systems and Applications, Ligano, Computing (L. Zadeh, J. Kacprzyk, D. Dimitrov, et al.,
Switzerland, 169-174. Eds.), Book Series: Artificial Intelligence Series – WSEAS,
18. Kim D.H., Park J.I. (2005) LNCS, 3645, Springer, WSEAS Press, 34-39.
Heidelberg, 366-375. 34. Roeva O. (2008) LNCS, 4818, Springer, Heidelberg, 601-
19. Kim J.S., Kim J.-H., Park J.-M., Park S.-M., Choe W.- 608.
Y., Heo H. (2008) World Academy of Science, Engineering 35. Roeva O., Tzonkov St. (2009) Int. J. Bioautomtion, 12,
and Technology, 47, 384-389. 1-12.
20. Kirkpatrick S., Gelatt C.D., Vecchi M.P. (1983) Science, 36. Roeva O., Pencheva T., Hitzmann B., Tzonkov St.
New Series, 220(4598), 671-680. (2004) Int. J. Bioautomation, 1, 30-41.
21. Krohling R.A., Rey J.P. (2001) IEEE Trans. Evol. 37. Wang Q., Spronck P., Tracht R. (2003) Machine Learning
Comput., 5, 78-82. and Cybernetics, 3, 1651-1656.
22. Kumar S.M.G., Rakesh B., Anantharaman N. (2010) 38. Yacoub A.H.Y., Buyamin S., Wahab N.A. (2011) Journal
International Journal of Computer Applications, 2, 1053- Technology, 54, 381-402.
1368. 39. Ziegler J.G., Nichols N.B. (1942) Trans. Amer. Soc.
23. Kumar S.M.G., Jain R., Anantharaman N., Mech. Eng., 64, 759-768.
Dharmalingam V., Begum S.K.M.M. (2008) Indian
Chemical Engineer, 50(3), 214-226.
24. Mahony T.O., Downing C.J., Fatla K. (2000) In: Process
Control and Instrumentation, University of Stracthclyde,
148-153.

Biotechnol. & Biotechnol. Eq. 26/2012/5 3277