Journal of Applied Logic 24 (2017) 97–108

Contents lists available at ScienceDirect

Journal of Applied Logic

www.elsevier.com/locate/jal

Sensorless control of variable speed induction motor drive using

RBF neural network
Pavel Brandstetter ∗ , Martin Kuchar
VSB – Technical University of Ostrava, Department of Electronics, 17. listopadu 15/2172,
70833 Ostrava-Poruba, Czechia

a r t i c l e i n f o a b s t r a c t

Article history: High power of modern digital signal processors and their decreasing prices enable
Available online 11 November 2016 practical implementation of different speed estimators which are used in the
sensorless control of AC drives. The paper describes application possibilities of
Keywords:
artificial neural networks for the sensorless speed control of the A.C. induction motor
Artificial neural network
RBF neural network drive. In the sensorless control structure of the A.C. drive, there is implemented
Vector control the speed estimator which uses two different artificial neural networks for speed
Sensorless control estimation. The first speed estimator uses a multilayer feedforward artificial neural
Induction motor network. Its properties are compared with the speed estimator using a radial
A.C. drive basis function neural network. The sensorless A.C. drive was simulated in program
Matlab-Simulink. The main goal of many simulations was finding suitable structure
of the artificial neural network with required number of neuron units which will
ensure good control characteristics and simultaneously will enable a practical
implementation of the artificial neural network in the digital signal processor control
system.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction

The evolution of A.C. variable-speed drive (VSD) technology was driven partly by the desire to emulate
the excellent performance of the D.C. motor, such as fast torque response and speed accuracy. Development
of modern VSDs is characterized by process made in various areas including electrical machines, power
electronics, control systems, and control algorithms [18,15,17,13,3].
At present, software tools can be used for evaluation of the rotor speed and position without usage of
the speed sensor. Due to modern digital signal processors it is possible to implement complex estimation
algorithm in real-time with the possibility of monitoring currents and voltages.

* Corresponding author.
E-mail address: pavel.brandstetter@vsb.cz (P. Brandstetter).

http://dx.doi.org/10.1016/j.jal.2016.11.017
1570-8683/© 2016 Elsevier B.V. All rights reserved.
98 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108

Fig. 1. Application of estimation techniques for sensorless electrical drive.

Removing rotor position sensor or mechanical speed sensor from the control structure of the electrical
drive leads to so-called sensorless electrical drive, which naturally requires other sensors for the monitoring
of stator currents and voltages (see Fig. 1).
In the scientific literature, there is possible to find many estimation techniques used for the sensorless
electrical drives which can be divided into following groups [10,11,9]:

• Estimators working in open loop with monitoring of stator currents and voltages.
• Systems with reference and adaptive models (MRAS).
• Estimators deriving benefits of space saturation of stator voltage utilizing phases voltage monitoring.
• Estimators exploiting possibilities of slot harmonic.
• Estimators with observers (Kalman filter, Luenberger observer, etc.).
• Estimator using injection methods.
• Estimators using soft computing methods.

The applications of the soft computing methods in the control of electrical drives, which include artificial
neural network (ANN) applications, discuss ANN based speed or position estimator applications. The ANN
speed estimators can lead to improved performance enhanced adaptive capabilities [4,5,7,6,14,8,12].
In the first part of the paper there is described a basic issue of the sensorless control of the vector-controlled
induction motor and its control structure. Next there are described speed estimators using the multilayer
feedforward (MLF) artificial neural network and the radial basis function (RBF) artificial neural network.

2. Speed control structure of induction motor

Modern control methods of the induction motor drives include the vector control of the induction motor
(IM). This control method uses a rotating coordinate system which is oriented on a magnetic flux vector
of the IM, for example on the rotor magnetic flux vector. In this rotating coordinate system, the controlled
quantities are direct quantities in the steady state.
The mathematical model of the IM in the rotor flux oriented coordinates [x, y] is described in the following
form:

diSx vSx dim

σTS + iSx = + ωim σTS iSy − (1 − σ)TS (1)
dt RS dt
diSy vSy
σTS + iSy = − ωim σTS iSx − (1 − σ)ωim TS im (2)
dt RS
 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108 99

Fig. 2. Speed control structure of induction motor drive with the ANN speed estimator.

dim
TR + im = iSx (3)
dt
3 Lm
TM = p im iSy (4)
2 1 + σR
dΩm
Jt = (TM − TL ) (5)
dt

where vSx , vSy – stator voltage components in [x, y] coordinate system; iSx , iSy – stator current components
in [x, y] coordinate system; im – magnetizing current; TS = LS /RS – stator time constant; TR = LR /RR
– rotor time constant; RS – stator resistance; RR – rotor resistance; LS – stator inductance; LR – rotor
inductance; Lm – main inductance; ωim – angular speed of rotating magnetizing current vector; ωR – rotor
angular speed; Ωm – mechanical angular speed; p – number of pole-pairs; TM – motor torque; TL – load
torque; Jt – total moment of inertia; σ – total leakage constant; σR – rotor leakage constant.
The vector controlled induction motor need two constants as input references (see Fig. 2): the torque
current component iSyref (aligned with the y-axis of the x, y coordinate system), and the magnetizing current
component iSxref (aligned with x-axis of the x, y coordinate system). The amplitude of the rotor flux is
determined by magnetizing current im and flux constant Kψ (ΨR = KΨ im ; where KΨ is flux constant). By
maintaining the amplitude of the rotor flux ΨR at a fixed value there is a linear relationship between torque
TM and the torque current component iSy (TM = KT ΨR iSy ; where KT is torque constant). We can then
control the torque by controlling the torque component of the stator current vector.
For variable-speed drive with the induction motor (IM), a cascade structure is often used. The cascade
control structure consists of several control loops, whereas the current control loops for the magnetizing and
torque current components iSx , iSy are subordinate and the flux and speed control loops are superior loops
(see Fig. 2).
Two motor phase currents are measured by current sensors and are transformed by the Clarke trans-
formation module. The outputs of this projection are current components defined in the stator coordinate
system iSα and iSβ . These two current components are the inputs of the Park transformation that gives
the current components in the rotating reference frame [x, y]. The iSx and iSy current components are
compared with the references iSxref and iSyref . As induction motor needs the rotor flux creation in order to
operate, the reference magnetizing current imref must not be zero. The torque command iSyref is the output
of the speed controller. The outputs of the current controllers are reference voltage components vSxref and
vSyref ; they are applied to the inverse Park transformation. The outputs of this projection are reference
voltage components vSαref , vSβref , which are the components of the stator voltage vector in stationary
stator reference frame [α, β]. These are the inputs of the vector PWM (Block of Pulse Width Modulation).
100 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108

The speed controller determines the reference value of the torque current component iSyref and the flux
controller determines the reference value of magnetizing current component iSxref . The parameters of the
classical PI controllers have been set by various mathematical and experimental methods. Modern design
approaches include soft computing methods, such as application of neural networks, fuzzy logic, genetic
algorithms, etc.
For vector rotation of variables from stator coordinate system [α, β] to rotating coordinate system [x, y],
an orienting angle γ is necessary. To obtain its value a block of estimation of orienting quantities (BEOQ)
is used which is based on a current model of the induction motor. In the block BEOQ there is calculated
magnetizing current im and orienting angle γ using stator current components iSα , iSβ and rotor angle ε.
The vector control is very sensitive to variations in the rotor time constant TR. Adaptation of this rotor
time constant TR is thus required, and it is necessary to estimate this parameter in order to maintain it
equal to its rated value.

3. ANN speed estimation of induction motor

In general, artificial intelligence could be used in almost all control parts of electric drives. Especially in:
identification, estimation of state parameters, control and diagnostics. Their usage could lead to the useful
improvement of necessary characteristics of the controlled electrical drives [1].

3.1. Speed estimator using multilayer feedforward ANN

For the speed estimation of the induction motor, there is necessary to find out the proper neural network
architecture. Up to now there does not exist any widely used rules to choose the proper network architecture,
neither the number of neural units. There are just generally accepted rules as a simplest network with highest
accuracy.
An appropriate structure of the neural network with suitable input variables, which will implement the
views defined by the following equation:

ωm(k) = f [vSxref (k) , vSxref (k−1) , vSyref (k) , vSyref (k−1) , iSx(k) , iSx(k−1) , iSy(k) , iSy(k−1) , w] (6)

where f is the activation function and w is a vector of weighting and threshold coefficients.
First it is necessary to design right structure of the artificial neural network and it is also important to
determine such inputs to ANN, which are available in structure of vector control and from which is able to
estimate a rotor speed of the induction motor. A recommended method for determination of ANN structure
does not exist, so the final ANN was designed by means of trial and error. The main goal was to find the
simplest neural network with good accuracy of speed estimation. This is the key for industry use of ANNs.
For the MLF ANN speed estimator (see Fig. 3), it was tested various structures of the artificial neural
network for different speed areas, for example 8-10-1, 8-20-1, 8-22-1, 8-26-1. The simulation results of many
structures were not so good, especially estimated signal (output of the MLF ANN speed estimator) contained
higher ripple.
Finally, a three layer ANN 8-28-1 was used which contains 28 neurons in the hidden layer with tanh
activation function, and one neuron in output layer with linear activation function (see Fig. 3). The ANN
has 8 inputs for quantities vSx(k) , vSx(k−1) , vSy(k) , vSy(k−1) , iSx(k) , iSx(k−1) , iSy(k) , iSy(k−1) , (voltage and
current components of stator space vectors vS , iS of the IM) and 1 output ωm(k) (rotor angular speed).
The ANN speed estimator has been implemented into speed control structure of induction motor drive with
vector control (see Fig. 2). The A.C. drive was simulated in program Matlab. Training stage is performed
in Matlab using Levenberg–Marquardt algorithm [1].
 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108 101

Fig. 3. Structure of the MLF ANN speed estimator.

3.2. Speed estimator using radial basis function ANN

The Radial Basis Function (RBF) neural networks belong to forward neural networks with one hidden
layer containing local units. The RBF network is a type of ANN for applications to solve problems of
supervised learning e.g. regression, classification and time series prediction [1,2,16].
The radial basis functions are powerful techniques which are built into a distance criterion with a respect
to the centre. That means that their characteristic feature is that their response decreases (or increases)
monotonically with distance from a central point. The centre, the distance scale, and the precise shape of
the radial function are parameters of the model, all fixed if it is linear. The RBF network is nonlinear if the
basis function can move or change size or if there is more than one layer of hidden units.
The RBF neural network is three-layer neural network. The input layer is used for transfer of the input
values. A second layer, marked as a hidden layer, consists of RBF units. These units realize individual radial
basic functions. The third layer contains the output unit with the linear function.
The parameters of the RBF neural network, such as number of layers and neurons, depend on the
character of the solved problem. In the case that a small number of neurons are used, the RBF neural
network will not be able to cover all dependences on training data. The ANN learning time is depending
on the ANN structure. If the number of neurons is large, then the learning time increases and the excessive
amount of training data network causes poor generalization ability.
Output of the RBF network is defined as follows:


N
 
y= φi wi , φi = φ x − ci  (7)
i=1

where wi is weight of ith centre ci , φi is some radial function and x − ci  denotes distance between x
and ci .
The radial basic function belongs to a special class of mathematical functions. These functions monoton-
ically decrease or increase with increasing distance from the centre. The most popular Gaussian activation
function is defined as follows:
 
z2
φ(z) = exp − 2 , z = x − ci  (8)
2r

where r is the radius of the radial base function.

This paper is focused on single-layer networks with functions which are fixed in position and size. Nonlin-
ear optimization is used, but only for regularization parameters in ridge regression and the optimal subset
of basis function in forward selection. It is also important to realize, that if we stay in the world of linear
algebra, analysis and computation are easier and quicker [16].
102 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108

Fig. 4. Structure of the RBF ANN speed estimator.

Generally, the hidden unit function is a Gaussian function which, in the case of scalar input, is:

 
(x − c)2
h(x) = exp − (9)
r2

Its parameters are its centre c and its radius r. A Gaussian RBF monotonically decreases with distance
from centre. In contrast, a multi-quadratic RBF which in case of scalar input is monotonically increased
with distance from centre [16].
The second type of the speed estimator was designed with the RBF ANN (see Fig. 4). The number of
radial units is very high, approximately about two hundred units.
In the first step, it is necessary to design the suitable structure of the RBF neural network. It is also
important to choose such inputs which are available in the speed control structure. Fig. 4 shows the RBF
architecture with the appropriate input variables. There are three layers: the input layer, the hidden layer
with the RBF units and the output linear layer. The RBF neural network has 8 inputs for quantities vSx(k) ,
vSx(k−1) , vSy(k) , vSy(k−1) , iSx(k) , iSx(k−1) , iSy(k) , iSy(k−1) , (voltage and current components of stator space
vectors vS , iS of the IM) and 1 output ωm(k) (rotor angular speed).
We have chosen the method of trial and error to find the simplest neural network with sufficient accuracy
of the speed estimation. A comparative procedure was created. The RBF neural network with the appro-
priate architecture was firstly realized. Then the RBF unit field of coverage was changed from the radius
value of one. In fact it means a more sporadic or denser layout of the RBF units, which is expressed by a
lower or higher number of RBF units. The input training data set for the RBF ANN, which was obtained
from the speed control structure, is shown in Fig. 6 and Fig. 12.
In the first RBF neural network, 165 RBF units were used. The time response was good and the difference
of the real and estimated speed was satisfactory. In the next RBF neural network, the denser layout of
330 RBF units and the output signal was better and the difference of the real and estimated speed was
lower. We also tested the thin layout of 55 RBF units and the response was not satisfactory. The disadvantage
of the structure with the denser layout is that it requires a higher memory and computation time.
The proper training RBF algorithm can be performed in the MATLAB-Simulink software. Within the
research work, the whole simulation validity process of RBF network training was established. The training
algorithm examination was performed for the speed estimator with RBF neural network and it was found
that the forward subset selection training algorithm was the most suitable choice for this topic. Also, the
RBF neural network architecture was checked and as noted above the architecture was decidedly the best
option.
 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108 103

Fig. 5. RBF parameters verification in MATLAB.

The speed estimator was implemented into the speed control structure as mentioned above. To check the
RBF network behavior there was uploaded different training data and checking program was written in the
Simulink (Fig. 5).
The RBF parameters (it means C, R and weight matrixes w) were uploaded into this program and
various data input sets were sent into their inputs. Then was just checked the RBF output and concerned
if the training process was successful. If so, the RBF parameters were saved and then implemented into the
main speed control structure [16].
Radial Basis Function neural networks do not need any data standardization for their calculation process
as the feedforward neural networks (activation functions range is −1; 1). If the original acquired data
for the training process are used, the only difference is that the trained RBF network has a much higher
data range of the weight matrix parameters. The only problem with the not standardized input data is
that’s it comes into being a huge computational value range demand. That became a problem, especially
in consideration with the fix point architecture of the signal processor. Therefore, the training data was
according to the requirement and the neural network is then trained with this data. In other words, the
incoming and outgoing data from the RBF network must be standardized as well [16].

4. Simulation results

Design of the ANN speed estimators was carried out in Matlab-Simulink. The parameters of the IM are:
Pn = 2.7 kW, VSn = 380/220 V, ISn = 7.51 A, Ωn = 1360 rpm (ωn = 142.41 rad s−1 ), J = 0.02 kg m2 .
A voltage inverter, which is the most important part of the frequency converter, works with DC-link voltage
VDC = 650 V and uses comparative pulse-width modulation (PWM). Switching frequency is chosen fsw =
10 kHz.
The training data set for the low speed area with a load jump TL = 5 N m was obtained from simulated
A.C. drive with the induction motor (see Fig. 6 and Fig. 7). For this purpose 50,000 samples were recorded
for each of the input and output signals. It was achieved an error 5.10−5 during training stage.
For the control quality evaluation of the sensorless A.C. drive, it is important to assess the speed time
course in different situations. Thus, the time course of the reference rotor angular speed was defined. The
simulation was performed for the reference speeds which represent different speed areas.
104 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108

Fig. 6. Normalized input training data set for MLF and RBF ANN estimator (low speed area).

Fig. 7. Normalized output training data for MLF and RBF ANN estimator (low speed area).

Fig. 8. Reference and real rotor angular speed responses ωmref = f (t) [rad s−1 , s]; ωm = f (t) [rad s−1 , s] (low speed area).

For presentation in this paper, the reference speed is chosen which includes start up the induction motor
to 100 rpm (ωmref = 10.47 rad s−1 ) and the subsequent braking to 0 rpm (ωmref = 0 rad s−1 ). During this
operation the A.C. drive is loaded by load torque TL = 5 N m at time t = 0.15 s.
Fig. 8 shows time courses of the reference and real rotor angular speed of the vector controlled induction
motor. In Fig. 9 and Fig. 10, there are shown time courses of the reference and estimated rotor angular
speed which were obtained by the MLF ANN estimator and RBF ANN estimator. It is obvious that RBF
ANN estimator has better properties than the MLF ANN estimator (see Fig. 11).
The training data set for the high speed area without load was obtained from the speed control structure
of the A.C. induction motor drive (see Fig. 12 and Fig. 13). For this purpose 100,000 samples were recorded
for each of the input and output signals. It was achieved an error 3.10−5 during training stage.
For presentation of the A.C. drive activity in the high speed area, which includes de-excitation regime
of the induction motor, there is chosen the reference speed from zero speed to 2000 rpm (ωmref =
209.42 rad s−1 ) and the subsequent braking to zero speed (ωmref = 0 rad s−1 ). During this operation the
 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108 105

Fig. 9. Reference and estimated rotor angular speed responses of the MLF ANN estimator ωmref = f (t) [rad s−1 , s]; ωmest =
f (t) [rad s−1 , s] (low speed area).

Fig. 10. Reference and estimated rotor angular speed responses of the RBF ANN estimator ωmref = f (t) [rad s−1 , s]; ωmest =
f (t) [rad s−1 , s] (low speed area).

Fig. 11. Difference between real and estimated speed of the MLF and RBF ANN estimator Δωm = (ωm − ωmest ) = f (t) [rad s−1 , s]
(low speed area).

A.C. drive works without load. The mentioned de-excitation regime of the induction motor begins after
exceeding the rated speed Ωn = 1360 rpm (ωn = 142.41 rad s−1 ).
Fig. 14 shows time courses of the reference and real rotor angular speed of the vector controlled induction
motor. In Fig. 15 and Fig. 16, there are shown time courses of the reference and estimated rotor angular
speed which were obtained by the MLF ANN estimator and RBF ANN estimator in high speed area.
It is again obvious that RBF ANN estimator has better properties than the MLF ANN estimator (see
Fig. 17), but at the expense of higher requirements for the practical realization of the speed estimator.
106 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108

Fig. 12. Normalized input training data set for MLF and RBF ANN estimator (high speed area).

Fig. 13. Normalized output training data for MLF and RBF ANN estimator (high speed area).

Fig. 14. Reference and real rotor angular speed responses ωmref = f (t) [rad s−1 , s]; ωm = f (t) [rad s−1 , s] (high speed area).

5. Conclusion

The RBF neural network and its application for the speed estimation was chosen. In order to validate
the most interesting theoretical assumptions of RBF neural networks in the control of electrical drives,
simulation using Matlab-Simulink was carried out. The interesting simulations were realized which confirmed
the rightness of proposed structures and good behavior of developed speed estimators.
The different structures of the RBF neural network have been investigated. The structure with the denser
layout had the best behavior in comparison with the others structures. The whole implementation process
starting with the data acquisition and adjusting, RBF neural network training and the implementation into
the vector control was described.
The practical implementation of the sensorless control by the help of RBF neural network will require
the usage of the high performance DSP. The usage of the fixed-point processor makes the implementation of
the RBF neural networks complicated, and above all, the computation demand too high. The combination
 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108 107

Fig. 15. Reference and estimated rotor angular speed responses of the MLF ANN estimator ωmref = f (t) [rad s−1 , s]; ωmest =
f (t) [rad s−1 , s] (high speed area).

Fig. 16. Reference and estimated rotor angular speed responses of the RBF ANN estimator ωmref = f (t) [rad s−1 , s]; ωmest =
f (t) [rad s−1 , s] (high speed area).

Fig. 17. Difference between real and estimated speed of the MLF and RBF ANN estimator Δωm = (ωm − ωmest ) = f (t) [rad s−1 , s]
(high speed area).

of the high number of neural units and exponential function, make it possible for this application to be used
in the near future with the usage of the floating-point DSP.
A suitable topic for future research is to investigate whether an on-line network weights adjustment
can be incorporated into the proposed ANN speed estimators without impairing stability. This option is
available as a result of the linear output combination of the values from the hidden nonlinear layer and their
appropriate weights. It might also be possible to use the new floating-point digital signal processors for the
neural network application.
108 P. Brandstetter, M. Kuchar / Journal of Applied Logic 24 (2017) 97–108

Acknowledgements

The paper was supported by the projects: IT4Innovations Centre of Excellence project, reg. no.
CZ.1.05/1.1.00/02.0070 funded by the European Regional Development Fund and the national budget of
the Czech Republic, Center for Intelligent Drives and Advanced Machine Control (CIDAM) project, reg. no.
TE02000103 funded by the Technology Agency of the Czech Republic, project reg. no. SP2016/83 funded
by the Student Grant Competition of VSB-Technical University of Ostrava.

References

[1] P. Brandstetter, M. Kuchar, J. Friedrich, Application of RBF neural network in sensorless control of AC drive with
induction motor, in: International Joint Conference SOCO’14–CISIS’14–ICEUTE’14, in: Advances in Intelligent Systems
and Computing, vol. 299, Springer Int. Publishing, 2014, pp. 217–227.
[2] J. Cai, Z.Q. Deng, A RBF neural network based sensor less control scheme for switched reluctance motor, Int Rev.
Electrical Eng. 7 (6) (2012) 6026–6034.
[3] P. Chlebis, P. Vaculik, P. Moravcik, Z. Pfof, Direct torque control methods for three-level voltage inverter, in: 10th
International Scientific Conference on Electric Power Engineering 2009, 2009, pp. 352–356.
[4] E. Corchado, M. Wozniak, A. Abraham, A.C.P.L.F. de Carvalho, V. Snasel, Recent trends in intelligent data analysis,
Neurocomputing 126 (2014) 1–2, Special Issue SI.
[5] T.H. dos Santos, A. Goedtel, S.A. Oliveira da Silva, M. Suetake, Scalar control of an induction motor using a neural
sensorless technique, Electr. Power Syst. Res. 108 (2014) 322–330.
[6] M.R. Douiri, M. Cherkaoui, A. Essadki, Neuro-genetic observer speed for direct torque neuro-fuzzy control of induction
motor drive, J. Circuits Syst. Comput. 21 (7) (2012).
[7] F.F.M. El-Sousy, Adaptive dynamic sliding-mode control system using recurrent RBFN for high-performance induction
motor servo drive, IEEE Trans. Ind. Inform. 9 (4) (2013) 1922–1936.
[8] P. Fedor, D. Perdukova, Z. Ferkova, Optimal input vector based fuzzy controller rules design, in: International Joint Confer-
ence CISIS’12–ICEUTE’12–SOCO’12, in: Advances in Intelligent Systems and Computing, vol. 189, Springer International
Publishing, 2013, pp. 371–380.
[9] S.M. Gadoue, D. Giaouris, J.W. Finch, Sensorless control of induction motor drives at very low and zero speeds using
neural network flux observers, IEEE Trans. Ind. Electron. 56 (8) (2009) 3029–3039.
[10] J. Holtz, Sensorless control of induction motor drives, Proc. IEEE 90 (8) (2002) 1359–1394.
[11] Ch. Lascu, I. Boldea, F. Blaabjerg, Comparative study of adaptive and inherently sensorless observers for variable-speed
induction-motor drives, IEEE Trans. Ind. Electron. 53 (1) (2006) 57–65.
[12] F. Lima, W. Kaiser, I.N. da Silva, A.A. Oliveira, Speed neuro-fuzzy estimator applied to sensorless induction motor control,
IEEE Lat. Am. Trans. 10 (5) (2012) 2065–2073.
[13] I. Neborak, P. Simonik, L. Odlevak, Electric vehicle modelling and simulation, in: 14th International Scientific Conference
on Electric Power Engineering 2013, 2013, pp. 693–696.
[14] T. Orlowska-Kowalska, M. Dybkowski, Performance analysis of the sensorless adaptive sliding-mode neuro-fuzzy control
of the induction motor drive with MRAS-type speed estimator, Bull. Pol. Acad. Sci., Tech. Sci. 60 (1) (2012) 61–70.
[15] D. Perdukova, P. Fedor, Virtual laboratory for the study of technological process automation, Int. J. Eng. Educ. 29 (1)
(2013) 230–238.
[16] O. Skuta, Modified Concepts of the Artificial Neural Network Architecture in the Modern Control of Electrical Drives,
PhD. thesis, VSB-Technical University of Ostrava, 2008.
[17] V. Sladecek, P. Palacky, D. Slivka, M. Sobek, Influence of power semiconductor converters setup on the quality of electrical
energy from renewable sources, in: 11th International Scientific Conference on Electric Power Engineering 2010, 2010,
pp. 527–531.
[18] P. Vas, Artificial-Intelligence-Based Electrical Machines and Drives, Oxford Science Publication, 1999.