Comparison of the Performance and Energy

Consumption Index of Model-Based Controllers

Luis I. Minchala-Avila1 , Kenneth Palacio-Baus1 , Juan P. Ortiz2 , Juan D. Valladolid2 , and Juan Ortega3
1 Universidad de Cuenca, Department of Electrical & Electronic Engineering
Ave. 12 de Abril, Cuenca, Ecuador
2
Universidad Politecnica Salesiana, Department of Automotive Engeneering
Calle Vieja 12-30, Cuenca, Ecuador
3
Universidad Catolica de Cuenca, Control & Automation Research Group
Ave. de las Americas, Cuenca, Ecuador

AbstractThis paper presents a comparison of the perfor-
mance of different control algorithms in two types of systems;
one exhibiting fast dynamics and the other slow dynamics.
The first control system regulates the speed of a DC motor,
while the second control system regulates the temperature of an
electrical heater. This systems performance comparison pretends
to evaluate the energy consumption, as well as the controllers
transient response in order to identify the best control strategy for
each system. System models are obtained through the responses
to a pseudorandom binary signal (PRBS) and the least squares
fit method using an auto-regressive model with an exogenous
variable (ARX). The implemented control algorithms used in this
study are: pole placement regulator (statespace controller) with
integral error processing, auto-tunable proportional-integral-
derivative (PID) controller, neural PID controller, unconstrained
model predictive control (MPC), fuzzy PID controller, neuro
fuzzy controller, bayesian controller and an optimal quadratic
regulator (LQR). A detailed analysis of the performance and
energy consumption index is performed, that allow the cate-
gorization of the control strategies in accordance with their
performance.
Index Terms Digital controller, performance index, model,
temperature control, DC motor speed control
I. I NTRODUCTION
Digital controllers offer a wide range of flexibility for
automation engineers, such as easiness in reconfiguration, opti-
mal system behavior, etc. The selection of an adequate control
strategy for any application should consider characteristics
such as complexity, control effort, efficiency, tracking error
tolerance, robustness and adaptability [1]. The correct choice
of a controller represents a challenge for the designer, as this
decision determines the system efficiency.
Numerous speed control schemes for dc motors and temper-
ature control loops have been tested. Reference [2] presents
the results of a control of a DC motor by an adaptive neural al-
gorithm. The authors highlight the importance of reducing the
control structure complexity in order to reduce implementation
costs. Authors of [3] presents simulation results of the a DC
motor speed control, using PI, IP, fuzzy and neurofuzzy
controllers. System responses evaluation demonstrates that the
proposed IP controller has the best performance, since the PI
controller has an undesired overshoot, and the fuzzy controller
978-1-5090-1629-7/16/$31.00
2016
c IEEE
produce a sluggish response as well as the neurofuzzy driver.
The reduced performance of the fuzzy and neuro-fuzzy drivers
is caused by the lack of a systematic design methodology,
such that a trial-and-error design is required. Experiments and
simulation results presented in [4] and [5] demonstrates that a
fuzzy controller can be designed to obtain better performances
than PID, PI and IP controller. In [6] two approaches to self
tuning PID temperature controller are tested experimentally;
the technique is based on relay and the integral square time
error criterion set point. Although in both criteria the system
has a continuous steadystate error, the latter is more efficient.
There is a variety of performance indexes in assessing
controllers, e.g. in [7] the integral of the square error system
response is proposed as an indicator of performance of a
tracking strategy. In [8] the speed of loop control is chosen as a
performance index. These measurements are suitable whenever
a set number of samples are defined [9].
Comparisons among digital controllers for performance
measuring have already been made by previous researches,
reaching several conclusions: a tuned PI with small sampling
time generates better performance than an untuned driver with
dominant dead time [8]. Moreover, in [1] three types of fuzzy
controllers are compared through the performance measure of
a PI controller and a classic digital PID, concluding that the
fuzzy controllers are more robust, faster and more flexible than
traditional digital controllers.
From the design perspective, choosing the right control
strategy for different systems can be challenging, i.e. deciding
the best driver in energy expenditure. This paper experimen-
tally analyzes the measurement of two performance indicators,
tracking and energy consumption, for seven different control
strategies applied to two different systems with slow and fast
dynamics. From the comparison of experimental data, the
control strategy with the highest performance is defined for
each system.
II. S TATE - SPACE MODELS OF THE SYSTEMS
A. Speed control of a DC motor
Fig. 1 shows the electromechanical system of the DC
motor. The electromechanical system consists of the following
components:
 A DC motor which supports a control signal up to 5 V
corresponding to a maximum speed of 4000 rpm;
A plate attached to the motor shaft with 36 holes. The
plate is coupled to an encoder, which enables the speed
calculation;
A brake that allows to introduce disturbances to the
system.

Fig. 3: Scheme of pole placement regulator with error integra-

tion

algorithms. Statespace models (1) and (2) are used to calcu-

Fig. 1: Electromechanical system of the DC motor late feedback gains, integral gains, etc. Bayesian, neural PID
and neuro-fuzzy controllers are designed using process data
and control rules.
To find the plant model of each system, the classic least
square adjustment to an ARX model method is used. The
A. Pole placement regulator
identification signal used is a 7-bit PRBS varying between 0
and 5 V. Sampling time of the identification test is 50 ms for Fig. 3 shows the block diagram of this control strategy. The
the speed control, and 1 second for the temperature control. error integration of this controller allows a small steadystate
The statespace model of the dc motor is the following: error, which is desirable. Additionally, the systems model does
not have have zeros in z = 1, which is detrimental to track
[10], and that is compensated by this controller in order to
xk+1 = Gxk + Huk (1) track constant reference signals.
yk = Cxk The control action uk is defined by the ecuation (4) with vk ,
   
1.2447 1 26.29 the integrator, given as the Eq. (3). K1 and K2 are constant
xk+1 = xk + uk
3.448 0 31.58 gains, y(k) is the system response and r(k) is the reference
signal.
 
yk = 1 0 xk
B. Temperature control system
Fig. 2 shows the details of the temperature control system, vk = vk1 + rk yk (3)
which comprises an halogen bulb that heats the system and uk = K2 xk + K1 vk (4)
a temperature sensor. By using the same methodology as
before, the temperature control system is characterized by the Additionally, this control strategy requieres the design of an
following representation in state space: observer with the following representation:

x xk + Huk + KTe (yk C

k+1 = G xk ) (5)

B. Relay-based PID auto-tuning

The auto-tuning based on the relay method automates the
process of finding the ultimate gain and ultimate oscillation
period Ku and Tu (when the system reaches its critical point),
Fig. 2: Temperature control system by which, using the tuning Ziegler Nichols rules, the PID is
parametrized (Kc , TD and Ti ) [11].
Fig. 4 shows the scheme of the relaybased PID auto-
    tuning controller. If tuning is required, the PID controller
1.6545 0.6574 1 is disconnected and the feedback relay is activated. When a
xk+1 = xk + uk (2)
1 0 0 stable limit cycle (critical condition system) is set, the PID
parameters are calculated and the controller is reconnected to
 
yk = 0.0089 0.0023 xk
the process [6]. In this selftuning method, a procedure of
III. C ONTROL STRATEGIES measuring the noise level and the exchange time between two
Modelbased design is a mathematical method of address- reference levels calculated using the measured level, is used.
ing problems associated with the design of complex control The control law is given by Eq. (6).
 Fig. 4: Scheme of relay-base PID autotuning Fig. 5: Scheme of unconstrained MPC

uk = uk1 + w1 ek w2 ek1 + w3 ek2 (6)

 
Ts Td
w1 = Kc 1 + +
Ti T
  s
Td
w2 = Kc 1 +
Ts
Td Fig. 6: Scheme of the fuzzy PID controller
w3 = Kc
Ts
where, Ts is the sampling period; Kc , Ti y Td are the PID
controller parameters. With a constant reference in the optimization window, the
control law is obtained from the Eq. (9), with u(k) given
C. Neural PID controller by Eq. (10).
This controller is characterized, as all PID controllers, for
not requiring the plant model for the design. The tuning of uk = uk1 + uk (9)
the controller gains is automatic, and can be used even with uk = T
( + R) 1 T
[Rs Fx(ki )] (10)
nonlinear plants. The control law in Eq. (6) can be expressed as
the weighted sum of the error of current sample and previous where F and are matrices obtained from the recursive
samples, plus the previous control signal as expressed in Eq. process shown in Eq. (11), R is a Hermitian matrix and Rs
(7) [12], where w is the vector of weights from previous is the reference vector of length Np .
errors. The weights are updated according to Eq (8), where the
weights values are updated using a learning coefficient and
y(ki + 1|ki ) = CGx(ki ) + CHu(ki ) (11)
by the current error multiplied by previous samples error. The
 
weights, w1 w2 w3 are calculated through the Adaline y(ki + Np |ki ) = CGNp x(ki ) + CGNP 1 Hu(ki )
neuron training algorithm. + . . . + CGNp Nc Hu(ki (12)
+Nc 1)
 T Y = Fxk + uk (13)
uk = uk1 + wk1 ek ek1 ek2 (7)
 
wk = wk1 + ek ek ek1 ek2 (8) E. Fuzzy PID controller
In this design, fuzzy control methods for supervision and
D. Unconstrained MPC
assignment of the PID gains in each operating condition are
This control strategy is based on the system statespace used [13]. Fig. 6 shows the scheme of the fuzzy controller.
model. The general architecture of this controller is shown in The fuzzy controller serves as a supervisor, which is always
Fig. 5. The purpose of this controller is to calculate the value trying to recognize whether the PID is not properly tuned, and
of uk for optimizing the plant output yk in accordance with a then adjusts the gains according to a set of fuzzy rules to obtain
fixed trajectory rk . the best performance.
An horizon of control, as the sliding window of Nc samples,
is defined in which uk is fully described, although only the F. Neuro-fuzzy controller
first sample (u[1]) is used to be applied to the system. A The adaptive neurofuzzy inference system (ANFIS) con-
prediction horizon is also determined, Np , which is the number troller combines the techniques of artificial neural networks
of samples of the sliding window in which the state variables and fuzzy logic. According to [14] there are three models of
are predicted. Control criterion is further defined; a function neurofuzzy controllers. The model proposed in this work is
of cost to be minimized, which is generally the integral error the cooperative one, where the neural network determines the
obtained in the prediction window. parameters for the fuzzy system. Fig. 7 shows the scheme of
 Fig. 7: Scheme of neuro-fuzzy controller

the neurofuzzy controller, mainly composed by three blocks:

the fuzzy controller, neural network which adjusts the gain
parameter of the controller and the plant to be controlled. Fig. 8: Bayesian network
G. Bayesian control
The bayesian control is based in probabilities. It uses the
calculation of the probabilities from Bayes theorem, given by 4000
Eq. (14). 3500

P (B|A)P (A) 3000

P (A|B) = (14) Angular speed [rpm]
P (B)
2500
where P (A) is the probability of occurrence of the event A,
Integral
P (B) is the probability of occurrence of the event B, P (B|A) 2000
MPC
is the conditional probability, that the event B occurs since the LQR
1500
event A occurs, P (A|B) is the conditional probability that the Autotuning PID
Neural PID
event A occurs since the B event occurs. 1000
Fuzzy PID
The Bayesian control is described as the process of finding Bayesian
the most probable hypothesis scenario according to evidence. 500 NeuroFuzzy
Each hypothesis corresponds to a state of the system. Through Reference
0
the probability of occurrence of a state, a control action that is 0 0.5 1 1.5 2 2.5 3
Time [s]
responsible for making the change in the system is determined.
The process of finding all system probabilities is called in- Fig. 9: Step response of the DC motor speed control
ference mechanism. A general structure of a Bayesian network
is presented in Fig. 8, where ci are the causes nodes, fi : are
the effect nodes, ii : are the intermediate nodes, and X: is the
system state
1
For each network node a probability that forms a set of J = x (N )Sx(N ) + (17)
probabilities E = e with values of ci and fi is identified. The 2
1 N 1
goal is to know the probability given by Eq. (15). [x (k)Qx(k) + u (k)Ru(k)]
2 k=0
P (x|e) = P (X = x|E = e) (15) IV. E XPERIMENTAL RESULTS
Fig. 9 shows the DC motor angular speed transient response,
To train the Bayesian network controller, results of the using a step reference signal of 3000 rpm, for each controller
unconstrained MPC were used. design.
H. LQR controller Fig. 10 shows the temperature transient response, using a
step reference signal of 40 C, for each controller design.
This controller is also based on the statespace model of
In [10] the index tracking performance is defined as the
the plant. The control law is expressed in Eq. ( ref (16).
Eq. (18). According to this definition, the power of the error
signal is a measure of the ability of the controller for tracking
uk = Kxk (16) a reference. The lower the performance index, the better the
controller is.
where K is a gain resulting from solving the Ricati equation
applying Lagrange multipliers to solve the problem of mini-
mizing the control criterion of Eq. (17). J =
k=1 ky(k) r(k)k
2
(18)
 2
x10
46 60

44
50

42

Tracking performace index

Temperature[C]

40
40

38 Integral 30
MPC
36 LQR
Integral
Autotuning PID MPC
20
Neural PID LQR
34
Fuzzy PID Autotuning PID
Bayesian Neural PID
32 10
NeroFuzzy Fuzzy PID
Reference Bayesian
30 NeuroFuzzy
0 50 100 150 200 250 300 350 0
0 50 100 150 200 250 300 350
Time [s] Time [s]

Fig. 10: Step response of the temperature control

Fig. 12: Performance index for temperature control

7
x10
25 x10 7
Integral 120
MPC Integral
LQR MPC
20
Tracking performance index

Autotuning PID 100 LQR

Neural PID Autotuning PID
Energy consumption index

Fuzzy PID Neural PID

15 Bayesian 80 Fuzzy PID
NeuroFuzzy Bayesian
NeuroFuzzy
60
10

40
5
20

0
0 0.5 1 1.5
Time [s]
Fig. 11: Performance index for DC motor speed control
Figures 11 and 12 show the performance index of the DC
motor speed and the temperature control systems, calculated
according to Eq. (18), over time.
In [10] the performance index under control penalty is
further defined according to Eq. (19).
J = 2
k=0 ke(k)k + k=0 ku(k)k
2
This index can be considered as an energy consumption
measure of the controlled system, since it includes the energy
of the error signal and the control signal.
Figures 13 and 14 show the performance index under control
penalty for the DC motor speed control and the temperature
control systems.
In the speed control system, the LQR controller exhibits the
best performance. According to the performance factor under
control penalty graph (Fig. 11), the system consumes less
power with this controller. Additionally, its step response is
2 2.5 3 0
0 0.5 1 1.5 2 2.5 3
Time [s]
Fig. 13: Energy consumption index for DC motor speed
control
faster than the other controllers. The bayesian and neural PID
controllers provide sluggish response of the system, showing
low energy efficiencies.
In the temperature control system, the unconstrained MPC
has the best performance. The neurofuzzy controller provides
(19)
faster response and a better tracking capability of the reference.
The MPC control also shows the best energy efficiency.
Responses of neurofuzzy and PID neuronal controllers are
similar with respect to the step reference signal; however,
the neurofuzzy controller responds faster and consumes less
energy than the neural PID. The LQR controller is inadequate
in controlling this system. Despite its low power consumption,
the reference level is not successfully reached.
The PID controller with fuzzy supervision has the best
tuning in both systems. The graphs of the performance factor
under control penalty show that this tuning generates the
 x102 The PID controller, which does not require the model of
1400 the plant, can be tuned using different techniques. It has
Integral
MPC been shown that a fuzzy supervisory system provides the best
1200
LQR tuning, in both the fast and the slow system.
Autotuning PID
Energy consumption index

1000 Neural PID

Fuzzy PID
Bayesian
800 NeuroFuzzy
600
400
200
0 Conference on, vol. 2. IEEE, 2001, pp. 479483.
0 50 100 150
Time [s]
Fig. 14: Energy consumption index for temperature control
lowest energy consumption.
V. CONCLUSIONS
The performance of a control system cannot be completely
generalized as demonstrated in this research. It could be the
case that a control strategy shows a good performance in
systems with fast dynamics, while offering poor throughput
in sluggish systems.
The LQR controller offered a better efficiency in the speed
control system than the performance showed in the temper-
ature control system. The energy consumption was better in
this regulator in comparison with the control algorithms tested
in this research.
The MPC controller showed a better efficiency in the
temperature control system. The control signal of the MPC
is calculated iteratively in a receding horizon window, which
allows a low steadystate error. The higher the prediction hori-
zon, the better efficiency of this control algorithm; although,
it is a tradeoff of complexity and performance, due to the size
of the matrices involved in the calculations of the algorithm.
The bayesian controller depends on a careful adjustment of
the values of transition, probability of transition and changes
in the control variable in order to get a faster and more efficient
controller.
R EFERENCES
[1] M. S. Kalavathi and C. S. R. Reddy, Performance evaluation of classical
and fuzzy logic control techniques for brushless dc motor drive, in
Power Modulator and High Voltage Conference (IPMHVC), 2012 IEEE
International. IEEE, 2012, pp. 488491.
[2] C. Machbub, A. S. Prihatmanto, and Y. D. Cahaya, Design and
implementation of adaptive neural networks algorithm for dc motor
speed control system using simple microcontroller, in Power Electronics
and Drive Systems, 2001. Proceedings., 2001 4th IEEE International
200 250 300 350 [3] S. P. S.R. Khuntia, K.B. Mohanty and C. Ardil, A Comparative Study
of P-I , I-P , Fuzzy and Neuro-Fuzzy Controllers for Speed Control of
DC Motor Drive, vol. 3, no. 2, pp. 15, 2009.
[4] O. Montiel, R. Sepulveda, P. Melin, and O. Castillo, Performance of a
Simple Tuned Fuzzy Controller and a PID Controller on a DC Motor,
no. Foci, pp. 531537, 2007.
[5] S. C. Ms. Priya Yadav, Salma Khan, Yogesh Chauhan, Speed Control
of DC Motor, no. 10, pp. 17191724, 2014.
[6] E. D. Bolat, K. Erkan, and S. Postalcioglu, Experimental autotuning
pid control of temperature using microcontroller, in Computer as a
Tool, 2005. EUROCON 2005. The International Conference on, vol. 1.
IEEE, 2005, pp. 266269.
[7] J. Chen, S. Hara, L. Qiu, and R. H. Middleton, Best achievable tracking
performance in sampled-data systems via lti controllers, Automatic
Control, IEEE Transactions on, vol. 53, no. 11, pp. 24672479, 2008.
[8] J. Quellmalz, M. Rehm, H. Schlegel, and R. Neugebauer, Influence
analysis on the model comparison performance index for servo drive
control, in Mechatronics-Mechatronika (ME), 2014 16th International
Conference on. IEEE, 2014, pp. 242247.
[9] H. Fujioka, S. Hara, and H. Tokunaga, Robust performance synthesis
for sampled-data feedback control systems, in Decision and Control,
1995., Proceedings of the 34th IEEE Conference on, vol. 2. IEEE,
1995, pp. 17761781.
[10] T. Bakhtiar and S. Hara, Tracking performance limits for simo discrete-
time feedback control systems, in SICE annual conference, 2004, pp.
18251830.
[11] C.-C. Yu, Autotuning of PID controllers: A relay feedback approach.
Springer Science & Business Media, 2006.
[12] D. Uduehi, A. Ordys, and M. Grimble, Multivariable pid controller de-
sign using online generalised predictive control optimisation, in Control
Applications, 2002. Proceedings of the 2002 International Conference
on, vol. 1. IEEE, 2002, pp. 272277.
[13] J. Zhang, D. Yu, and S. Qi, Structural research of fuzzy pid controllers,
in Control and Automation, 2005. ICCA05. International Conference
on, vol. 2. IEEE, 2005, pp. 12481253.
[14] P. B. JUAN JOSE and R. RAFAEL, Diseno e implementacion neu-
rodifuso para control de un posicionador robotico, Ph.D. dissertation,
2012.