Article Not peer-reviewed version

A New Paradigm in AC Drive Control:

Data Driven Based Control by Learning
of the High Efficiency Data Set.
Generalizations and Applications to a
PMSM Drive Control System

Madalin Costin * and Ion Bivol

Posted Date: 7 October 2024

doi: 10.20944/preprints202410.0516.v1

Keywords: data driven control (DDC); model predictive control (MPC); permanent magnet synchronous
machine (PMSM); searching algorithm; radial basis function (RBF) neural networks (NN); interpolation
technique

Preprints.org is a free multidiscipline platform providing preprint service that

is dedicated to making early versions of research outputs permanently
available and citable. Preprints posted at Preprints.org appear in Web of
Science, Crossref, Google Scholar, Scilit, Europe PMC.

Copyright: This is an open access article distributed under the Creative Commons
Attribution License which permits unrestricted use, distribution, and reproduction in any
medium, provided the original work is properly cited.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

Disclaimer/Publisher’s Note: The statements, opinions, and data contained in all publications are solely those of the individual author(s) and
contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting
from any ideas, methods, instructions, or products referred to in the content.

Article

A New Paradigm in AC Drive Control: Data Driven

Based Control by Learning of the High Efficiency
Data Set. Generalizations and Applications to a
PMSM Drive Control System
Madalin Costin 1,2*, Ion Bivol 1
1 Department of Electrical Engineering and Energy Conversion Systems, “Dunarea de Jos” University of
Galati, 800008 Galati, Romania
2 Department of Automatic Control and Applied Informatics, “Gheorghe Asachi” Technical University of

Iasi, 700050 Iasi, Romania

* Correspondence: [email protected]

Abstract: This paper presents a new means to control the processes, which involve the electric energy
conversion. Electric machines fed by electronic converters provide the useful power defined by the inner
product of two generalized energetic variables: the effort and the speed. The novelty of the paper idea is to
control the desired energetics variables by a direct Data Driven Control (DDC) law in terms of the pair effort,
speed and corresponding voltage controls of the electronic converter. It must be mentioned that the same
desired useful power might be obtained with different controls at different efficiencies. The regularization
problem was solved by selecting from a knowledge database, the maximum efficiency operation points.
Knowing a reasonable number of optimal efficiency operation points, an interpolative Radial Base Function
(RBF) control was built up. The RBF algorithm can find by interpolation the optimal controls for any admissible
operation points of the process. The control scheme developed for this purpose has an inner DDC loop, that
perform the converter control based on measured speed and demanded torque by the outer loop which handles
the speed. A comparison of the DDC with the Model Predictive Control (MPC) schemas highlights the
advantages of the new proposed control.

Keywords: data driven control (DDC); model predictive control (MPC); permanent magnet
synchronous machine (PMSM); searching algorithm; radial basis function (RBF) neural networks
(NN); interpolation technique.

1. Introduction
Electric Drive Systems (EDSs) have getting a real interest in the least decades by developing
advanced technologies in digital processors, signals proceedings and power systems. The main
benefits provide by developing EDSs can be summarized as follows: fast transients, high power
density, large range of speed and power, low cost, safety operation, robustness, modular components
etc.
Increasing number of EDSs applications (see automotive) implies new control techniques to
improve the performances of the energy conversion process. Firstly, the efficiency improving of the
EDSs can be done from the conception phase, when there is selected the Permanent Magnet
Synchronous Machine (PMSM) that presents a high efficiency as a consequence of the using of
modern technologies of manufacturing of permanent magnets. Secondarily, the increasing of the
EDSs with PMSM may be obtained by using an adequate optimal control law designed by a certain
methodology.
The traditional control strategies are based on the physical equations of the process model, for
example the common used in EDSs are: Field Oriented Control (FOC) with the mathematical model
in an appropriate coordination frame which it is used by an adequate decoupled control changes for
torque and flux, the loops being usually designed with classical PI linear controllers [1,2], Direct

© 2024 by the author(s). Distributed under a Creative Commons CC BY license.

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

Torque Control (DTC) – a bang to bang control implemented by hysteresis controllers [2] and Model
Predictive control (MPC) which uses a future control sequence obtained by the process model
predictor in the aim to find the best control law sequence by minimising a knower objective function
[3].
The specific particularity of the advanced EDSs is that the input voltage and frequency are not
directly applied from the power grid, which presents fixed values. The command is obtained by using
a power interface represented by a power electronic invertor, which allows the giving variable values
of voltage and frequency for obtaining the imposed performances [4]. This task is crucial for met the
advanced results, which can improve the close loop performance of the EDSs [5,6].
In the least years, the MPC control strategy has been extensively developed and improved, being
an actual motivation for research and industry [7-11]
]]. More advanced studies are dedicated to improve the control results, taking into account the
nonlinear magnetic circuit [12], advanced optimization method [13], robustness [14] or fault tolerant
capability [15], respectively. However, MPC strategy is strongly based on the model of the process,
which in practice can differs from the one of the real system, leading to the worsening of the control
performances.
In EDSs, the Artificial Intelligence Methods (AIMs) are well approached on control system. Thus,
the torque minimization was successfully obtained by using Neural Network (NN) method [16], the
Radial Basis Function Proportional-Integral-Derivative (RBF-PID) controller developed in [17] leads
to improve the tracking results and an iterative learning control strategy reduce the torque ripple
[18]. More advanced AIMs can be founded in the survey [19].
The Reinforcement Learning (RL) is an advance popular framework used to control systems and
devices in a wide range of applications because of its ability to autonomously find control polices to
achieve a desired goal without assuming that the environment are known. Thus, in [20] it is proved
that the actual methods of deep RL leads to obtain a suitable control structure of the adaptive
controller. Further, a low-complexity gradient descent solution with backtracking iteration approach
for MPC Quadratic Programming (QP) problem leads by minimizing to reduce the number of the
searched control inputs and improve the control performances [21]. In spite of the various benefits
obtained by AIMs, these control advanced algorithms require a large amount of hardware and
software resources.
A new trend in the control theory applications is based on data-driven control (DDC) that
discusses the transition from Model-Based Control (MBC) applications to Model-Free Control (MFC)
adaptive structures, involving the learning in a given control framework context [22–25]. The DDC
method is an advanced strategy that can learn the dynamics [26], or can let to obtain different
formulations via regularizations and relaxations methods, bridging the direct and indirect data
driving control [27]. Indirect data driven control is formulated as bi-level optimization problem. That
is, first a model fitted from the data in the inner identification, before the model is used to control the
outer problem. Model is used for many reasons: first is the availability of powerful analysis and
design of the control methods. Another advantages is that the models are very compressed compared
to the data driven representation. At least, for more details about stabilization, optimality, and
robustness behavior of DDC see [28].
Also, it can be mention that there can be obtained predictive control applications, which are
directly obtained from the data set, which does not include all the main features of the classical
predictive control strategy, as: integrator that ensure a offset-free control, the constraint handling,
feed forward action, and disturbance control model, which make the strategy to be more versatile for
real-time applications [29]. Furthermore, an algorithm designed for control an unknown stochastic
linear time-invariant systems illustrate the high closed-loop performance obtained by data-enabled
predictive control strategy which has a novel robust distribution [30]. A simultaneous MPC and DDC
strategy with relaxed assumption on the model in a proper framework shows how a no convex
optimal problem is transformed into a one with convex constraints [31]. The input‐mapping based
data‐driven model predictive control is successful used for systems with bounded disturbance [32]
or for advance learning [33]. For generalization, in [34] is discussed a data-enabled predictive control
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

parametric algorithm that combined the identification method obtained by a learning procedure and
a control one of the dynamical systems, resulting in an equivalent MPC formulation, which is valid
for both linear time-invariant system, which is directly applied, and for nonlinear systems, that
requires a regularization procedure.
As expected, an emerging technology on control system development may be based on both
DDC and MPC strategies. Although, EDSs with represent a rapid process for control applications,
with the help of the advance sensors and adequate numerical signal processing there can be collected
and processed a large amount of data that is sufficient for implementing the DDC solution. Thus,
based on White-Box (WB) modeling estimation by least squares technique, in [35] a DDC-MPC
method is proposed for current controlling of PMSM, that leads to compensate the modeling errors
and, to improve the control performance for both dynamic and steady-state regimes. In relatively
recent paper [36,37], it has been shown the transition from the classical model-based approach
towards data-driven optimal control strategies, which take into account the model predictive control
paradigm, leading to develop the data-enabled predictive control. At least, in the reference [38] it is
shown that a fractional PI controller for a PMSM drive can be obtained by a DDC algorithm in certain
practical conditions.
MPC strategy and then recently starting by DDC have done many successfully solved
problems. An important issue in the EDSs with PMSM is related on the efficiency improving by
control means. In literature [39–41] can be founded various solution for improving the efficiency of
EDSs with PMSM. Most of them are based on the modeling of the power losses with algebraic
relation, taking into account some specific constants of materials. The minimize of an analytic cost
function with constraints will give the optimal solution that ensures a high efficiency. This approach
has an elegant mathematical description, but also many limitations: the constants of the materials
depend of the magnetic/electric fields, there are losses that cannot be modeled analytically (abnormal
losses) and many methods are designed for partial losses.
Finally, there is an extensive literature focusing on the MFC strategy (philosophy). Research
started early in 1956 through the project [42] “A proposal for the Dartmouth Summer Research Project on
Artificial Intelligence” by J. McCarthy, Dartmouth College, M. L. Minsky, Harvard University, N.
Rochester, IBM Corporation, C. E. Shannon, Bell Telephone Corporation. It is concluded in this research
project: ‘’if a machine can do a job, then an automatic calculator can be programmed to simulate the machine’’.
Paraphrasing we can say mutatis mutandis: if a classic automatic control based on a physical model can
control a process, then an informatic system can control the same process without a physical model. This is the
main philosophy followed in the paper to show how a MBC strategy offers good enough possibilities
to develop a MFC in certain energetic constraint. The addresing paper made a transion from the MBC
to MFC in the DDC paradigme for high performance objective.
For a large set of the phisical processes the main controlled quantities are effort and flow. Their
product represents the useful power. In the paper we develop a new technique to find adequate
controls of the effort and flow in terms of maximum efficiency, that is the minimum losses. These
appropriates controls are extracted from an aprioric DataBase (DB) in terms of required effort and
measured speed. The aprioric DB contains mainly the steady state measurements in the open loop
tests for: the speed, the effort and the corresponding controls. An RBF interpolativ system is trained
so that, based on an few set of the optimal experimental data, to exctract the controls for any desired
speed and flow. In the on line operation of the process the initial DB is adaptive. An on line
optimisation system search the new optomal controls when the process parameters are changed. DB
are free from process parameters and in consecace from the process model.
The paper shows that an informatic system based on an experimental database, without
knowing the parameters of the process, can do the same job as a MPC control system with multiple
advantages. Thus, we propose a new combined linear PID and DDC, where the control variables are
directly obtained from an experimental database by a RBF methodology. The key of DDC is to find
the relation of the energetic variables with the control variables, by a RBF -NN interpolative surfaces.
These surfaces are built with a set of the optimal experimental support points extracted from a large
knowledge database.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

In the following, we demonstrate that the proposed combined linear PID and the RBF-NN
methodology can successfully dynamically simulate EDSs with PMSM, controlled by the MPC
methodology, without knowing any parameters of the motor or driven process.
Respecting the state of art, in the paper there are proposed the following new contributions:
(i) A new combined linear and DDC control methodology that uses the process control variables
obtained from an experimental database built by a RBF methodology.
(ii) In the case of the EDSs, the combined control methodology is in terms of an outer time domain
PID control of speed and an inner implicit RBF control loop.
(iii) The inner control loop is based on an experimental database, which contains the optimal controls
corresponding to any speed and torque of process operation points. The control law is in implicit
form that is a set a steady state process operation points.
(iv) The RBF methodology compresses the set of all admissible operation points by network
architecture, with a few optimal data basis functions.
(v) We present a novel control scheme in which the demanded effort by the PI controller and
measured speed are the input for the RBF inner loop. The output of the inner loop is the
inverter’s modulator input, which select the proper inverter’s voltage vector sequence of the
invertor.
(vi) The inner RBF control loop is adaptive, that is the inverter control is temperature dependent
because the parameters of the process changes with motor is temperature.
The organization of this paper is structured as follows. Thus, in the Section II is presented how
is obtained a data free control algorithm based on the physical modeling of the processes. The section
III is dedicated to DDC of AC drives, where the illustrative application of PMSM drive control is
done. The comparative results obtained by MPC and DDC techniques are presented in Section IV. At
last, in Section V there were summarized the main benefits obtained by the proposed novelty control
method.

2. Problem Statement. From Classical Physical Modeling of Energy Conversion Flow to Data
Free Model Control

2.1. Problem Statement

The energy conversion flow is an essential objective for ensuring of the social and economic
stability. In industry and residence, many applications are designed to deal with the macroscopic
energy conversion that implies three main types of fields: electromagnetics, mechanics and
thermodynamics (Figure 1).

Electromagneti
cs

Thermodynami
Mechanics
cs
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

Figure 1. The diagram of the macroscopic energy flow conversion.

Starting from these energy conversion possibilities, in practice there can be developed different
conversion processes subject of adequate constraints.
In the various physical processes like electrical, mechanical, hydro- or thermal, the desired
useful work is universally characterized by a pair of energetic variables: the effort e (t) and flow s (t).
The inner product of these generalized variables is the instantaneous useful power:

p2 (t) = e(t)  s(t) . (1)

The different forms of primary power p1(t), can be converted in useful forms by various means.
Let be η the efficiency of the power conversion process. Then, primary power p1(t) is easily obtained:

p1 (t) = p2 (t) /  . (2)

If, for example, we refer to a mechanical process of motion which develop a useful force fu and
a linear velocity v, the motion being described in external (cartesian) coordinates, then we have e = fu,
and s = v. If the motion is rated to the drive motor shaft, then e = m and s = ω, where m is the developed
torque and ω the angular speed. In the case of an electrical process, the electrical voltage represented
in the complex domain, e=u and the corresponding current s= i. Similarly, in the case of Hydro,
Ventilation, Air Conditioning (HVAC) processes, the pressure e=P and the flow rate s=Q.
Usually, the above energetic macroscopic process are well-modeled by the actual intrinsic
physical theories. In the control theory, the using of this model with lumped parameters have some
vulnerabilities: high sensitivity to parameter variations, incomplete modeling of some loss
components, linearity on a small range, low control accuracy at high frequencies etc. However, the
analytic modeling has the advantage of simplicity, being in many approaches a starting point in the
developing of advanced control methods.
The transition from MBC to model MFC is illustrated in Figure 2. For this aim it is considerate a
combined PI and inner control (cascaded or not) for a given process described by adequate physical
model equations.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

MBC
u
sref eref s
PI Process
- MFC

a) The basis control scheme.

eref-MBC MBC
sref-MBC uMBC sMBC
PI Process
- MFC

b) Stored data from MPC running.

dMBC MFC u*MFC

c) Designing of MFC strategy: trainning phase

MBC
sMFC
sref-MFC eref-MFC uMFC
PI Process
-
MFC

d) Designing of MFC strategy: testing phase

Figure 2. The transion from MBC to MFC.

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

The inner control is switched from MBC to MFC whenever is necessary, giving on the output
the control u as is suggested in Figure 2(a).
In the first stage exposed in Figure 2(b), the inner control is switched to MBC. At this step, there
selected a collection of the main data in steady-steate regime that coresponds to the high efficiency,
grouped in:
T
d MBC = uMBC y MBC e MBC sMBC  , (3)

where has been considerate that in steady-state regime we have eref=e.

Now, the next step is the designing of the MFC controller. This controller has a interpolative-
adaptive structure with a vraiable structure, being able to learn from MBC data, having as input the
data stored (3) in a sufficient number, and the output u*MFC as is shown in Figure 2(c). The MFC
algorithm is trainned in a open structure, and has the ability to learn from the experince given by
MBC
Having closed the trainning phase of MFC, the testing phase in done in the close loop structure
as is indicated in Figure 2(d).
The transition from MBC to MFC depends of the process that is subject of control strategy. This
procedure may be applied to the process with complex, nonlinear and multivariable structure.

2.2. The DDC Strategy of the Energy Conversion Process

The MFC tehnique represents a modern approach with a large applicability in practice. From
MFC control techniques, a particular direction are described by DDC strategy which is developed in
the following.
The numerical processing of the signals that refers to the training phase are carried out with the
MBC strategy, which is why the MBC acronym will be ignored in the symbolization of signals,
distinguishing between the acronyms DDC and MPC whenever is necessary.
The state model of this nonlinear deterministic conversion processes is described by:
 x = f( x(t ), u(t ), θ(t )) .
 (4)
 y = g( x , u )

The state x, control input u and output y, respectively, are restricted to the sets of adequate
dimensions: x(t ) X  , u(t )  U  and y(t ) Y 
N M P
. Further, in the vector θ(t) there are
comprised the parameter of the process. The nonilinearty of the state model (3) is emphases by the
nonlinear functions f and g.
Let us to adopt the fallowing outputs variables:
T
y =  e s p1 p2  .
(5)
h( x , u) = p1

By usually discretizing techniques, the corresponding discrete state space model of the plant (4)
becomes:

 x( k ) = f( x( k − 1), u( k − 1, θ, Ts ) + ξ ,
 (6)
 y( k ) = g( x( k ), u( k ))


where the discrete states x(k), control inputs u(k), and outputs y(k) are matrices of appropriate sizes.
Also, in (6) ξ is the discrete noise of the process, while Ts represents the sampling time period.
The control problem to be solved is definited according to: seek for a control u(t) that will bring
a desired useful work:
t1
E2 (t ) =  p2 (t )dt , (7)
t0

when the input of the system is suppyled by the energy:

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

t1
E1 (t ) =  p1 (t )dt , (8)
t0

with a natural condition resulted from (2):

E1 (t ) = E2 (t ) / 
 E2 (t )  E1 (t ) , (9)
 1

such as for any other control uˆ(t ) , which also bring the same work E2, the energy E1 is larger:
t1 t1
t0
ˆ , x)dt   h( u , x)dt .
h( u
t0
(10)

In other words, we assume that for a given desired pair of the generalized variables e(t) and s(t),
there exist a bounded control u(t) and a bounded cost criterion, which can be espressed either in a
contionous set formulation:

Сc = 
t0
t1 t1
ε 2 dt ;  h( u , x )dt ; THDe E2 (t ) = cst. ,
t0  (11)

or in a discrete representation:
N
 N 

С d =  ε 2 ( k )Ts ;  h( k )Ts ; THDe E2 ( k ) = cst. , (12)

 k = 1 k = 1 


where there has been used the mean square error: ε = (e − e) + (s − s) .

2 2 2 ref ref

Also, in the cost formulation (11)-(12) there is used the Total Harmonic Distorsion THD factor
that is a measure of harmonic distortion, computed by:


E
n= 2
2
n , RMS
(13)
THDe = ,
E f , RMS

where E f , RMS and En , RMS represent the Root Mean Square (RMS) values of fundamental and
harmonics components of the effort e, respectively.
For the same desired work (7) the final cost 𝐶 which is lower, has better control u. The interval
(t1-t0) length is a time window with the required energy E2 (6), where t1/0 represents the end/start time.
It can be mention that the input power may be either positive p1  0 or negative p1  0 , as the
power is delivered or received by the process.
The first term from (11)-(12) refers at the control quality in the control horizon N length and the
second terms, respectively, refers to the input energy.
The numerical computation has a large aplicalbility in data computation. For our approach we
define cost – solving problem based on (12) as follows:

Сmin = arg min Сd

u ,x

 E = cst. , (14)
s.t. :  2
 ( k ) = max

where the constraints E2 and η ensure the uniqueness of the optimal cost problem.
The process model (4) is in time domain, having an explicit form. A more convenient approach
is developed in the following.
We introduce the implicit control law:

G( y , u) = 0 .
 T (15)
y =  e s 

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

In the model (15) the ouput of the process y was reduce in terms of the generalized variables
pair (e,s), and the corresponding input is u . In consequence, the output variables p1, p2 are free, being
unconstrained, which shows the control law G is not unique. A regularization procedure is
effectuated by selecting from the set (e,s) corresponding to the maximum efficiency power operation
points.
The command law of the process results from the explicit model (15):
 u = F( e , s) .
 −1 (16)
 F=G

Let consider a time window [t0 , t1 ] for requested output trajectories of the both generalized
variables:

 g1 (t ) = e(t )
 , t  [t0 , t1 ] . (17)
 g 2 (t ) = s(t )


By eliminate the time from relations (16) we have the implicit representation:

f (e , s) = 0 . (18)

Each discrete point kTs , with k=1,...,Nobs, where Nobs is the number of observations points, on the
trajectory (17) in the generalized variables (e, s) plane, has the input controls which represents the
supports points for the control surfaces:
u( k ) = f( e( k ), s( k ))
u = [u1 u2 ] . (19)
f = [ f1 f2 ]

The surfaces (19) in the continuous domain represent the endless set of steady state operation
points. Any stable operating point (e,s) is correlated with the corresponding’s controls (u1,u2). A
trajectory of motion (16) in the permissible range, beginning in a stable initial state (e0,s0) and reaching
the final state (ef,sf), passes through an countless set of intermediate energy states, since in classical
physics it is admitted that states of physical processes can vary only continuously. By using digital
process control techniques, the set of states travelled from the initial state to the final state becomes
countable due to sampling operations. An exhaustive condition of stability of process controlled by
the two fundamental energy quantities, effort e and velocity (flow) s is as follows: if the speed control
loop is stable in the sense that the required effort eref is on the control surfaces (19), the operating
points on the motion path will be stable. Through the sampling process, stability is reduced, the
longer the sampling period.
Generalizing we can say that the multivariable control law (19) can be generated with the
required precision, by means of a known set MF of operating points of the process:

MF = [y u]T . (20)

Through an interpolation algorithm AF associated with the MF database, for any current value
y  Y , a value u  U is generated which approximates the control law F( y) relation (15), so that:
u = AF (y) . (21)

The control law is determined by means of an experimental data set MF and some continuity
and smoothing properties of the surfaces (15), which are specific properties of energetic inertial
processes.
In summary, the control law has the form:
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

10

u = AF ( y )
, (22
u , y  MF

with properties:
u( k) = AF ( y( k)), ( u( k), y( k))  MF
. (23
k = 1, 2,..., Nobs

The RBF-NN is largely used in control system

]]. A typical RBF-NN structure used for obtaining the input plant command is depicted in Figure
3.

Figure 3. The RBF-NN used for obtaining the plant input.

The RBF-NN is used for two objectives, as training and testing:

Training: AF = RBF(MF )
Testing: u RBF ( k) = RBF(AF ,( e( k), s( k))) . (24)
⎯⎯
Input
→ ⎯⎯⎯⎯
Output
→

where by RBF has been denoted the function of the neural network.
Having a classical structure, the RBF is organized in three levels of layers as input, hiden and
output, respectively. Moreover, the weights obtained in the testing step are grouped in:

AF = [w1 w2 ... wn−1 wn ]T . (25)

The output of the RBF-NN is given by a linear combination:

u RBF ( k) = AFT φ = w11 + w22 + ... + wnn , (26)

where the activation function is given by Gaussian distribution
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

11

2
I l − l
−
l = e  l2
, l = 1, n . (27)

having the parameters σ and μ which are the basis width vector and center points of Gaussian
function, respectively.
Also, the input Gaussian function is represented by set:

I = {AF , e ref ( k), s( k)} . (28)

Enhancing the RBF depicted in Figure 3, in Figure 4 is illustrated the combined control structure
of a generic process, that involves an outer explicit flow control loop and a inner multivariable data
driven control loop.
The fundamental operation key of the DDC schema is the knowlage database. This database
containes a large set of process operation points, from which are limited steady-state ones. In the
block scheme from Figure 3, the actual values dMBC (k) of control, output, effort and flow is comprised
in the set:
T
d MBC ( k) = u( k) y( k) e( k) s( k) , (29)

where y(k) is the output of the process, and N is the number of colected data.
The next task is to select a suficient number of steady-state points d MBC ( k) from all the operation
points from dMBC (k) ,
T
d MBC ( k ) =  u ( k) y  ( k) e  ( k) s ( k)  ,
d MBC ( k )  d MBC ( k)
  MBC (30)
 d ( k) .
with:   0( k) ,
 t
0( k ) = 0( k ) 0( k ) 0( k) 0( k) T
  
 MBC
For obtaining the suficient number of steady-state points contained in the set d ( k) , the
operator used for data extraction will be symbolized by “°”.
Supplimentary, as input data, there are considered the aprioric informations included in the set:

Kdb = {RN ; EN ; HN } , (31)

where: rated specification of the process are contained in the set R N , envirovment and design
requirments which are contained in the sets EN and H N , respectively.
The input control data set will be computed based on both aprioric information and process
data. In fact, the design of the required database is quite complex, and will be presented in the next
section for the case of a concrete application. However, stating from the inputs of knowledges KDB

and data d ( k) the output of the database is

DF = W {Kdb ; d MBC ( k)} , (32)

where by W has been represented a numerical operator used for data processing in the
database.
The proposed PI and combinated DCC structure is illustrated ub Figure 4.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

12

Figure 4. A combined control structure with linear PI outer control loop and DDC inner control.

The trainning phase designing of DDC strategy is depicted in Figure 4(a), where are highlited
the main steps of DDC algorithm, while the testing phase made in close loop structure is represented
in Figure 4(b).
The DDC control structure from Figure 4(b) is desined for tracking the output of the process
s (k) to its reference sref-DDC(k) by using a linear PI controller, having the output eref-DDC(k) bounded
DDC

into the limits -/+emax.

The control data set MF contains the maximum efficiency operation points selected from the
knowledge database. The interpolation algorithm AF find the actual control u(k) based on demanded
effort e(k)ref and the measured flow s(k).
We assume that all points MF (19), also referred as nodes, or “support points” are distinct and
not collinear. Due to the intrinsic nature of the energy conversion process, the DDC law (15) is
geometrically represented as continuous differentiable surfaces which allows to find an
interpolation algorithm AF such that these surfaces passes exactly through each experimental
points MF (18). The maximum efficiency operation points (e,s) is selected from the knoulagy
database by a grid (direct) search method. These methodology involve setting up of grids in the input
space (e,s) and evaluating the power efficiency of each grid point. The operation point that
corresponds to the maimum efficiency was considered the best solution.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

13

Two interpolation methods that are widely used in different application fields (in areas such as
computer graphics, physical modeling, geographic information systems, medical imaging, and
more), have been adopted. We used two basic methods: Delaunay triangulation and related methods
and RBF-NN interpolation.
Let be a set of data and we wont to find a rule which allows us to deduce information about the
process we are studying at locations differet from those at wich we are obtained our measurements.
The interpolated value of a point, other than support points, was obtained by local interpolation
techniques with the three nearby points. One widely used approach is Delaunay triangulation of data
and the Voronoi diagram of a set of points, which is the dual of the first approach (Delaunay
triangulation).
The RBF network has its origin in performing the interpolation of a set of data points in a
multidimensional compact domain to an arbitrary accuracy, given sufficient number of data points.
It has a network architecture with weighed basis functions. The solution is optimal by minimizing a
functional containing a regularization terms. The RBF method compresses a very large volume of
data, with the help of a much smaller number of weighted basic functions in the training process. In
the exam (assessing) stage, the RBF network returns any point in the field covered by the data used
in the training stage by an interpolation process of the weighted basis function. The discrete model
of the control law is practically non-inertial, because the control u(k) at the current step k is obtained
by accessing the fast interpolation algorithm AF :

u( k ) = AF ( y( k )),
k = 1, 2,..., N , (33)
N = Td / Ts ,

where Td is the desired final time of the motion trajectories.

The new control methodology of the energy conversion process in terms of the effort and flow
was represented in Figure 4. The objectives (10)-(11) are accomplished by two control loops: a MIMO
inner control loop in terms of desired effort eref(t) and measured flow s(t) and an outer classical PI
speed control loop.

3. The DDC of AC drive. Illustrative Case Study: The DDC of a PMSM Drive System
The control method proposed in the privious section has the ability to offers a new alternative
of the classical methods which are based on the model accuracy. The novel approach is model free,
being suitable for a large class of applications.
In the following, we begin with the principle of the DDC scheme applied to AC drives.
Afterwards, there is done an illustrative application for a PMSM driving systems controlled
according to the methodology developed in this paper. As MBC method has been selected the MPC
strategy due to benefits of providing an optimal solution without any frequency modulator.

3.1. The DDC Principle of AC Drives

AC drives are founded in a large area of applications due to their height-performances resulted
in dynamic regime.
Usually, the cascade control structure of AC drives is used to separate the different dynamics of
electrical and mechanical parts, respectively, and then to have an independent control of each loop.
In addition to the advance control law, an important attention is paid for the electronic power
inverter used to obtain the voltage supply. This aspect will be largely approached in the following
sections.
This section is dedicated to the DDC principle applied to AC drives.
As mention in the second section, the DDC method is based on a learning technique. For our
approach, the learning is based on the collected data from MPC strategy running.
The stages on DDC design are illustrated in Figure 5.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

14

It is starting from a cascade structure, having the well-known quantities: speed (actual and
reference), torque (reference), voltage and current (components). A brief description of each stage can
be done as follows:
a. Stage I: Stored data from MPC cascade control running. In this stage there are selected a limited
data obtained from MPC cascade running in steady-state regime for: speed, torque, voltage and
current. Further, from this there is imposed the constraint of high efficiency, resulting in the
stored data;
b. Stage II: Training of DDC strategy. The stored data obtained on the previous stage serves as a
base in this training stage for DDC design which is done in a open structure. Usually, the stored
data are numerically processed in the order to improve the main features: efficiency, robustness,
tracking and disturbance rejection. This objectives are done by using a DB which offers multiple
possibilities of data processing;
c. Stage III: Testing of DDC method. Learning from the previous stage, the DDC strategy is able to
operate in any point required by the application of AC drive. The DDC method are implemented
with no cascade control structure as is shown in the figure.
The DDC strategy may be a emergence solution in the actual AC drives whose model parameters
sensitivity is still reported as an open problem.
Nevertheless, an important objective it represented by the improving of the efficiency in various
scenarios.
Usually, the electrical parameters of the model of AC drives are subject of variations as
dependence of exogenous conditions that can differ with environment conditions: temperature,
external fields, altitude etc. In this can there must be studied the robustness of the DDC strategy
under these circumstances.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

15

Selected Data Stored Data

Voltage
Speed Torque ° ¤
Ref. Ref. Speed
MPC Current
PIω Control AC Motor ° ¤
- Curent
° ¤
Torque

° High Efficiecy ¤
Constraint

a) Stage I: Stored data from MPC cascade control running.

Training: Stored Data

(RBF-NN)
DDC Control Output Data :Evaluation
Testing: Check Data

(Learning)

b) Stage II: Training of testing the RBF-NN DDC control

Speed Torque
Ref. Ref. Voltage Speed
DDC
PIω Control AC Motor

c) Stage III: Testing of the DDC control

Figure 5. The principle of the DDC strategy applied to an AC motor.

3.2. DDC of the PMSM Drive System

The PMSM drives are largely used in the adjustable applications that require high performance
control performances and high efficiency. A conventional structure used in practice includes a two
level power electronic inverter as in Figure 6. On this input of the power electronic inverter is usually
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

16

ensured a constant DC bus voltage UDC which is provided by a DC bus circuit that contains a DC link
voltage and a voltage rectifier. The main feature of this power structure is that the discrete logic
commnad Sabc={Sa,Sb,Sc} applied on the top switches {1,3,5} is negated applied to the bottom switches
{4,6,2}.

Figure 6. Power structure of the PMSM drive. .

The power inverter discrete command is provided by the control structure that is developed in
the following.

3.2.1. Plant Modeling: PMSM and Power Inverter

Naturally, the model of PMSM is developed in the three-phase coodinates (a,b,c). This model has
a complex structure, being difficult to be used in the practical close-loop control applications. A first
approach to avoid this drawback is to use ortogonal fixed coordinate (α,β) that offer some
improvments. At last, the ortogonal transfomation in a moving system reference frame (d,q) provides
the highest benefits of the model of PMSM for close loop applications. Indeed, the usual FOC
tehnique appied to PMSM is done by align the space vector of PM flux  PM with a north pole of the
rotor of PMSM.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

17

ωe
d
iq

id
α
PM
0

Figure 7. Space vector diagrame of a PMSM with rotor system frame orientation .

In addition, the stator current components (id,iq) of space vector current i admits to be colinear
and perpendicular with the PM flux direction. The both space vectors as PM flux  PM and current i
rotates with the electrical speed ωe.
For this kind of application, the pair of flow and effort mention above in the privious section it
corresponds to torque and angular speed:

(e , s) (m, m ) . (34)

The PMSM model obtained by FOC principle is composed by the currents and the motion
equations [8]:

did (t ) 1
= (ud (t ) − Rs id (t ) + Lqe (t )iq (t ))
dt Ld
diq (t )
, (35)
1
= (uq (t ) − Rs iq (t ) − Lde (t )id (t ) − e (t ) PM )
dt Lq

dm (t) 1  f 
=  m(t) − m (t) − m (t)  , (36)
dt J  zp 

where: (ud,uq),(id,iq) and (Ld,Lq) are voltages, currents and inductances, all expressed in (d,q)
coordinates, (ωe, ωm) are electrical and mechanical angular speed, respectively, Rs denoted the stator
resistance, ψPM represents the permanent magnet flux, m and mℓ are the electromagnetic and load
torques, respectively, and the group of constants represented by zp,J and μf that are poles pairs, total
inertia of drive and friction factor, respectively.
The system (35) – (36) was completed by introducing the expressions of the electromagnetic
torque, magnetic flux components, input and output power, respectively, as follows:

3
m(t) = z ( L − Lq )id (t )iq (t ) + PM iq (t ) , (37)
2 p d

 d (t ) = Ld id (t ) +  PM
, (38)
 q (t ) = Lq iq (t )
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

18

 ,
p1 (t) = Re{u i * } = Re (ud (t) + juq (t))(id (t ) − jiq (t ))
(39)
p2 (t) = m(t)  m (t)

where u and i are the phasors of voltage and current, respectively, i* represents the complex-
conjugate phasor of current, and by Re{◦} is denoted the real part of the applied quantity.
The equivalent circuit scheme of the PMSM based on the dq current model (35) is depicted in
Figure 8, where has been introduced the total ElectroMotive Force (EMF) force components:

ed (t ) = − Lqe (t )iq (t )
. (40)
eq (t ) = Lde (t )id (t ) + e (t ) PM

id(t) Rs Ld iq(t) Rs Lq

+ +
ud(t) uq(t)
ed(t) eq(t)
- -

(a) (b)

Figure 8. The dq equivalent circuit scheme of PMSM.

The Equations (34) – (38) describe the so-called dq model of the PMSM drive system. Based on
the dq model there is designed the control structure in a propper maner. However, the physical
operation of PMSM is done in the natural three-phase system reference frame. To obatin the desired
quantities there are used the Clarke and Park direct/indirect transformations, respectively, of the
quantity λ=u,i,ψ, as follows:

λ = T λ abc

−1 , (41)
λ abc = T λ

λ dq = Tdq λ
, (42)
λ = Tdq−1 λ dq

where vector-matrix quantities used are:

T
λ abc = a b c  ,
T
λ = d q  , , (43)
T
λ dq = d q  .

and the involved matrix-tranformations are given by:

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

19

 1 1 
1 − −  1 0 1 
2 2 2 2 
T =   ; T =
−1
3 ;
3 3 3 3 0 − 1
0 2 − 2   2 2 
  2   2   , (44)
cos  m cos   m −  cos   m +   cos  m sin  m 1 
2  3   3   −1 2  
Tdq = ; Tdq =   2   2  3 .
3  2   2   3 cos   m −  sin   m − 
 sin  m sin   m −  sin   m +   2 
  3   3 
  3   3 

The power invertor command will be supplied by the finite set values Sabc which contains the
optimal logical values requited for obtaining inverter control.
Based on finite set values Sabc the model of the two-level power inverter provides the three-phase
voltages:
1
u abc = U DC TI Sabc , (45)
3
where the power inverter matrix model is:
 2 −1 −1
 
TI =  −1 2 −1 (46
 −1 −1 2 
.

3.2.2. MPC with Finite Set

The DDC method is developed by using the main features of the MPC strategy that provides a
command law which corresponds to voltage of the inverter, computed by a switching functions set.
In this strategy, the main variables as current, voltage and electromagnetic torque are discretized on
every time instant Ts by forward Euler forward method, that gives to a generic variable z(t) the next
fist-time derivative linear approximation:
dz(t ) z( k + 1) − z( k ) ,
 (47)
dt z = z ( k + 1)
Ts

Taking into account Euler forward approximation method, the first-time derivate of current (35)
becomes:

 TR  Lq T
id ( k + 1) =  1 − s s  id ( k ) + Tse ( k )iq ( k ) + s ud ( k )
 Ld  Ld Ld
, (48)
 TR  Ld Ts Ts
iq ( k + 1) =  1 − s s
 i ( k ) − Tse ( k )id ( k ) − e ( k ) PM + uq ( k )
 Lq  q Lq Lq Lq

Now, from the first-time derivate of current (37), the predictive electromagnetic torque results:

3
m( k + 1) = zp ( Ld − Lq )id ( k + 1)iq ( k + 1) + PM iq ( k + 1) , (49)
2
The one-step-ahead MPC law is an optimal solving problem defined as follows:
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

20

g * = min gi
Sabc

  m − m( k + 1) + id ( k ),
ref
m  mrated
 g =  . (50)
s.t.  i  mref − m( k + 1) + emf rated − ( + L i ( k )) , m  mrated
  PM d d

i = 1  8; k = 1  N

where emfrated is the rated back motion EMF computed by:

emf = kem . (51)

with the back EMF constant:

ke = z p PM , (52)
the index i corresponds to the combination of switching state Sabc (see below the Table 1).

Table 1. The voltage space vector according to their corresponding switching function.

Index i Sabc Voltage space vector

1
(0,0,0)
v1 = 0

2 v 2 = 2 3U DC
(1,0,0)

3 v3 = 1 3U DC + j 3U DC
(1,1,0)

4 v4 = −1 3U DC + j 3U DC
(0,1,0)

5 v5 = − 2 3U DC
(0,1,1)

6 v6 = −1 3U DC − j 3U DC
(0,0,1)

7 v7 = 1 3U DC + j 3U DC
(0,1,0)

8 v8 = 0
(1,0,1)

A better ilustration of the voltage space vector of the power inverter and their hexagonal bounds
is done in Figure 9.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

21

Figure 9. The voltage space vector according to their coresponding switching states.

In the dq system reference frame, the components of the space vector v for the eight possible
states are obtained in a vector-matrix representation:

U
v d = 0 2 1 −1 2 −1 1 0  DC
3
. (53)
3
v q = 0 0 1 −1 0 −1 −1 0  U
3 DC
Further, in Figure 10 is presented the cascade control structure, with a PI speed control in an
outer loop and a MPC multivariable inner current control loop for the electronic power converter
which supply the PMSM. The PI speed control is designed via the classical pole placement method
[45].
The electrical energy provided from the power grid is of fixed values of voltage and frequency.
By and adevate rectifer there is obained a DC voltage. The DC-bus voltage UDC is filtred by a passive
Low Pass Filter (LPF) represented by a high value capacitor C, obtaining a constant value of the DC-
bus voltage UDC=cst.
The PI controller is used for trackig the mechanical speed ωmMPC its reference ωmref-MPC. For the
MPC formulation of the inner loop, there are measured the three-phase current (a,b,c), and then by
the Park/Clarke transformations are successively obtained the dq MPC current idqMPC(k)=[idMPC(k)
iqMPC(k)]T . Further, based on (48) is obtained the current prediction idqMPC(k+1)=[idMPC(k+1) iqMPC(k+1)]T.
Now,the MPC problem defined according to (50) will give optimal switching combinations SabcMPC.
For the future computations, there is additionally calculated the voltages udqMPC(k)=[udMPC(k) uqMPC(k)]T
based on a switch-voltage transformation S/u.
It can be mentioned that the mecanical speed ωmMPC is not directly measured. Firstly, there is
measured the postion of the rotor αmMPC(k) by an encoder (E), and then by derivation is obtained the
actual mechanical speed ωmMPC(k).
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

22

Power
Grid

C
UDC
-/+mmax
ωmref-MPC(k)
mref-MPC(k) MPC Current
PI SabcMPC
Contrroller
- idqMPC(k+1)
(50)

udqMPC(k) S/u

Delay Compesation idq (k)

MPC iαβ(k) iabc(k)

(Current Predictor) αβ/dq abc/αβ

(49)

PMSM
3~

αmMPC(k) mMPC

ωmMPC(k) mℓ
E

Figure 10. The MPC control PMSM drive.

3.2.3. DDC Law of PMSM

The main idea of DDC strategy is to learn from a well-known MBC method whose performances
has been already proved. In this sense, the DDC method of a PMSM drive is developed by selecting
a DB containing the data provided by MPC law. Following the metodology developed for DDC in
the sens of the maximum energy conversion process objective, for PMSM the output is given by speed
and torque y( k)  mMPC ( k) mMPC ( k) , while the control by dq voltage u( k )  u dq
MPC
( k ) . In this
conext, the training sequence becomes:
T
dMPC ( k) = udqMPC ( k) mMPC ( k) mMPC ( k) , (54)
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

23

Also, for the PMSM drive the training matrix of the RBF-NN is composed by speed and dq
volage of the reduced model (high efficiency points extraction):

MF = [y u]T ,
y = [mMPC ( k ) m MPC ( k )]T , (55)
u=u MPC
dq
( k ),

The training phase of the DDC controller is depicted in Figure 11. As mention above, the entire
metodology is based on the the learning of the experience of the high-value efficiency steady-state
data provided by MPC strategy.

Figure 11. Training phase of the DDC controller of PMSM.

The set of aprioric informations KDB includes rated data of PMSM, power inverter load and
other components, envirovment and design requirments of EDSs.
Once training is done, the DDC control structure may be done as in Figure 12. The output of the
DDC
DDC controller is given by the dq voltage u dq ( k ) . The optimal set of the switching functions
DDC
combinations S abc is then obtained by a command shaping block.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

24

Figure 12. The DDC control sctructure of a PMSM drive.

To have a succinct description, in Figure 13 is depicted a schematic representation of the DDC

control structure. The main controllers as PI and DDC are highlighted. The DDC algorithm is
represented in a sequential structure, where the command shaping has an internal computing loop.
The DDC algorithm has a complex structure. A first stage is the searching on the DB the
− DDC
corresponding command u refdq
. Once finded, an internal computing programming loop is designed.
− DDC opt − DDC
Further, there are compared the command u ref
dq
( k ) and the feedback v dq ( k ) (that is also an
optimal value) , resulting an error which is then integrated obtaining the the voltage control v cdq− DDC ( k )
that acts as input in the searching block. Finally, there is obtained switching functions develivred by
DDC
DDC method S abc that will feed the gate of the inverter.
More detailes will presented later in the pseudocode format of the DDC method.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

25

Figure 13. A schematic representation of the DDC control of the PMSM drive system.

3.2.4. Design of the DDC Database

The input data in DB is initially done from the MPC operation, selected from the high efficiency
values. After that, the data are processed according to various criteria. Hence, the acronym MPC will
be substituted, making the distinction between MPC and DDC whenever is necessary.
The operation of a PMSM drive depends on the required mechanical speed domain limits.
Generally, there are possible three situation that show how torque and outer power varies vs.
mechanical speed:

 m = cst.
I. m  (0, m ,b )  m  Constant torque
  P2

 m
m = II. m  (m ,b , m ,max )  m  Constant power , (56)
  P2 = cst.

III.    m
 m  Critical operation
 m m ,max
 P2

where: ωm,b is the based mechanical speed which corresponds to the rated one (ωm,b=ωmN), and
ωm,max represent the maximal mechanical speed.
In Figure 14 are illustrated the torque – speed characteristics of a PMSM drive system. Until base
speed ωm < ωm,b, the operation of PMSM is done on constant torque in region I. In the region II, where
the power becomes constant, there is used field-weakening technique. At last, the critical operation
done for ωm < ωm,max must be avoided (region III) for ensuring a safety operation.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

26

m,P2

II. Constant Power

III.
Critical
Operatio
n

I. Constant
Torque

0 ωmb ωm,max ωm

Figure 14. The torque – speed characteristics under iso-efficiency curves of a PMSM drive.

Let be Ω,M, P and H the sets of sets of torque, power, and their corrresponding efficiency:

Ω = {m j | j = 1, N } , (57)

M = {mj | j = 1, Nm } , (58)

P = {P2 j | j = 1, Np } , (59)

H = { j | j = 1, Nh } . (60)
where Nω/m/p/h is the number of points of each set.
As can be easlier oserver from Figure 14, some of the different paris of speed-torque (ωmj,mj) or
speed – power (ωmj,P2j) corresponds to the same value of efficiency. That means that the solution is
not unique.
Considering that all the sets (57) – (60) has the same number of points N = N p = N h = N , there
are introduced the next sets obtained by cartesian product:

Λm = {mj  mj | j = 1, N} , (61)

Λ p = {mj  P2 j | j = 1, N} . (62)
Now, there can be definited the following multi-valued maps:

fm : Λm → H , (63)

fp : Λ p → H , (64)
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

27

whose points are contained in the sets:

Γm = {(Λm ( j), j )| j = 1, N}, Γm  Λm  H , (65)

Γ p = {(Λp ( j), j )| j = 1, N}, Γ p  Λp  H . (66)

To avoid the efficiency multi-valued representation, we select the high efficiency points that
corresponds to each point, defining the functions:

fm* : Λ m → Hmax , (67)

fp* : Λ p → H max , (68)

where the maximal efficiency vector is :

Hmax = [1max 2max 3max Nmax ]T , (69)

and the contained sets are given by:

Γ*m = {(Λm ( j), jmax )| j = 1, N}, Γ*m  Λm  Hmax , (70)

Γ*p = {(Λp ( j), jmax )| j = 1, N}, Γ*p  Λp  Hmax . (71)

An illustrative representation of the transformation of te multi-valued map of the efficiency vs
torque/power into a single-valued functions of a PMSM drive characteristics is depicted in Figure 15.

fm/p f*m/p
Λ Η Λ Ηmax

Λ(1) η1 Λ(1) η1max

Λ(2) η2 Λ(2) η2max

Λ(3) Λ(3)
η3 η3max
…

…
…

η=ηmax

Figure 15. Transformation of the multi-valued efficiency map into a single-valued function.

An essential objective is the design of the DB. Generally, there are few steps that must be
followed in the order to obtain high computation accuracy as it is suggested in Figure 16.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

28

Figure 16. The database grid steps design of close loop applications.

Succinctly, the main followed steps can be described as:

I. Storage grid of high efficiency data. At this step there are collected the main data set in the
steady-state regime and knowledge which corresponds to a high efficiency value stored in
DF = W {Kdb ; d ( k)}

T
d ( k ) =  udq ( k ) idq ( k ) m ( k )   m ( k ) ,
. (72)
s.t.  ( k )=max

II. Database grid learning. The database sets are extended by a learning process, adding new
points contained in the set:
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

29

de ( k)  d ( k)
. (73)
s.t.  ( k )= max

III. Application to data grid searching. It is checked that the added new points corresponds to the
real operation for open loop applications.
IV. Close loop data grid control. Finally, the designed database grid is introduced in the desired
close loop application from Figure 16.
These steps can be changed/improved depending on the desired objectives. For example,
exogenous quantities may have a high influence on the system which are not on thermodynamic
equilibrium.
Numerical computations of the quantities delivered by DB are affected by errors and noises. The
DDC strategy is a searching discrete – interpolative method of the best solution within learning data
bounds.
In the order to improve the acuarcy of data processing there is used a moving window filter,
whose output is given by:
Lw −1
y [n] =
ftr 1
Lw u
k =0
nftr
[n − k ] , (74)

where Lw is the filter length, n is the current sample, and the input samples (nonfiltered) and output
samples (filtred), respectively, are:
T
unftr = unftr
dq
( k) mnftr ( k) mnftr ( k) , (75)

T
y ftr = udqftr ( k) m ftr ( k) mftr ( k) , (76)
T
unftr
dq
/ ftr
( k) = udnftr / ftr ( k) uqnftr / ftr (k)
where .
The moving window filter structure is represented in Figure 17. The input samples are introducd
by using the unit-delay operator Δ. On the output is obtained an average value of the moving
samples.
As can be observed from (75) – (76) the filtred qunatities are voltage, torque and mechanical
speed. Based on this, further there result the filering of both input and output powers of PMSM.
Finally, the filtred powers are used to compute the efficincy, that do not contains high numerical
noises and errors.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

30

Lw -1
Input
Samples
unftr(k)
Δ Δ … Δ

1/L
w Σ Σ Σ Σ

Averag
e
Sample
yftr(k)
Σ

Figure 17. The moving window filter structure.

3.2.5. DDC Matlab Implementation

The DDC of the PMSM is implemented in Matlab simulation environment, tacking into account
the main programming features of this software pakage.
To obtain a high analysis, in the paper is made up a comparasion between the MPC method and
DDC strategy of the PMSM drive.

MPC Algorithm
The basic idea of the model predictive control for the PMSM fed by the simplest inverter with
two voltage levels, having eight available state voltage switching vectors vi (Table 1) is the following:
the torque and speed can be controlled by direct selecting a proper switching invertor’s voltage vector
based on the machine parametric predictive model and an appropriate cost function g.
The objective of the Algorithm 1 represented in pseudocode format from Figure 12 is to give to
reader the necessary tools to understand the simulations results and to replicate the implementation
of the exposed algorithms thinking in Matlab environment.
The switching vector vopt , that was predicted to provide the most favourable process behaviour,
i. e, the one that minimize the objective function g, is considered to be optimal and is generated by
the inverter, applying directly the corresponding switch functions svs=[ S a, Sb, Sc ].
The parameters of the motor and the driven process are included in the vector em

(electromagnetic parameters). After initialisations of the finite states for the swithching vector (lines
1,2) and the rated values and state variables, the main iteration loop are opend (line 8).
The model requires the actual state variables x = [id iq m ] to be recalled from the previous
sampling period state x old = [idold iqold mold ] .
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

31

Algorithm 1. The MPC main program

% Initialization for the 8 possible values of the inverter voltages
1. vd=[0;2;1;-1;2;-1;1;0]*Udc/3; % Udc is DC-bus voltage [V]
2. vq=[0;0;1;-1;0;-1;-1;0]* sqrt(3)/3* Udc;
3. svs=[000; 100; 110; 010; 011; 001; 101; 111]; % Switching vector states inverter’s gate control
4. Pem=[mℓ, J, μ, Rs, Ld, Lq, ψPM]; % Declare process and machine parameters
5. ωref= ωrated; mmax=2.5*mrated; % Initialization of the desired speed and maximum torque
6. ωpold= ωint; idpold=0; iqpold=0; % Initialization of the state’s variables of the process
7. Ts=1e-4; Td=10s; N=Td/Ts; %Sampling period (Ts), Desired simulation time (Td)
8. for n=1:N
% Call function PI_control
9. [mref]=PI_control(ωref ,ωold ,mmax, Kp, Ki );
%The prediction-optimisation procedure
10. for i=1:8
% Call Model Predictive Control-(MPC_model) ( idpold,iqpold are the measured currents)
11. [id(i),iq(i),m(i),p1(i)]=MPC_PMSM_model(vd(i),vq(i),idpold,iqpold,,Pem,Ts); %Predictor
12. if(ωp<=(ωrated );
ref
13. g(i)=abs(m -m(i))+id(i);
14. else
15. g(i)=abs(mref-m(i))+abs(emfrated(i)-(ψPM+Ld*id(i))); % emfrated is rated bac-EMF
16. end % end if
17. % Search vdopt, vqopt such as g|min
18. [vdopt, vqopt, svsopt]=MPC_Command_Shaping(g(mref,m(i), emfrated ,p1(i), vd(i), vq(i)];
19. end % end for
% Call Process_Model
20. [id,iq,m,ω,p1,p2]= Process_PMSM_Model(vdopt, vqopt, idpold, iqpold, ωold , Pem);
% Update the states
21. idpold=idp; iqpold=iqp; ωoldp= ωold;
% Call Mouving_Average
22. [udftr,uqftr]=Mouving_Average(vdopt,vqopt,L,Ts); % L is the length of the moving window filter
% Save process outputs in Data_Base_1
23. [idd,iqq,mm,ωω,uddftr,uqqftr,pp1,pp2]=Data_Base_1(id(n),iq(n),m(n),ω(n),udftr,uqftr,p1(n),p2(n));
24. end % end for

Figure 18. The pseudocode format of the MPC algorithme of a PMSM drive control.

The electrical and mechanical parameters respectively, e and m , of the plant are grouped in
the electromechanically set of parameters em , as follows:
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

32

em
(t ) == [ Rs Ld Lq  PM m J f ] = e m
(77)
e m
.
The MPC algorithm starts with the initializing the main data used in the computing loop.
A first programing syntax is developed for implement the PI speed controller:

 mref  = PI _ control ( ref ,  old mmax , K p , Ki )

Input
(78)
Output

where  ref / old the reference /actual mechanical speed is, mmax denotes the maximal value of the
electromagnetic torque, and Kp/i are the parameters of the PI controller. The PI_control generates the
required value of torque reference (set point) mref.
The relation (78) describes a typical programing syntax that contain the Input/Output sets of
quantities. This type of programing syntax structure was fallowed in the rest of the paper.
The MPC strategy takes as input the voltages vd/q, the previous measured currents idold/ q , the
electromecanical set of parameters (t ) and sampling time instant of MPC calculated around ten
em

times higher then the simulation discretizing one:

idp ( i ) , iqp ( i ) , m p ( i ) , p1 ( i )  = MPC _ PMSM _ model (vd ( i ) , vq ( i ) , idold , iqold , em (t ),10  Ts ) . (79)
Thinking in the Matlab environment implementation, the function which simulate the PMSM
model has as the input the controls (ud,uq) - the components of the supply voltage, the old states, the
load torque and the electomagnetique parameters and as the outputs - the current components (id,iq),
the electomagnetique torque m and the angular velocity ωm:

[id , iq , m, m ] = Plant_PMSM_model [udold , uqold , idold , iqold , mold , m , em (t ), Ts ] .

Output Input
(80)

Finally, the programs were grouped and compacted in a software application of the proposed
control structure.
The equations of the physical model (80) are in the continuous time domain. The variables of the
model take values in the real domain . In fact the PMSM is supplied by the power electronic
inverter which provide in the simplest case, a eight finite switching state vector:

v dq = {vd (i), vq (i )}, i = 1,...,8 . (81)

The components (ud, uq) takes values on the set vdq. For this reason, these quantities must be
shaped. According to the principle algorithm of the power inverter command, within each sampling
period Ts, the components (vdopt,vqopt) and the duration d (0 <d> Ts) are calculated so that the machine
currents resulting from switching supply voltage v dq opt
= [vdopt , vqopt ] has to be as close as possible to the
ideal sinusoidal form of the three-phase currents:
opt
Find : v dq until  THDiabc = min
, (82)
which admits the syntax:

[vdopt , vqopt ] = Inverter_Command_Shaping ud ( k), uq ( k), Ts  . (83)

The command algorithm used by MPC control is implicit, that is different from classical version.
Within each sampling period Ts, the command algorithm looks for the optimal components (vdopt,vqopt),
which minimizes the g criterion function:

 (
[vdopt , vqopt ] = MPC_Command_Shaping  g vd ( k), vq ( k), m, mref , id ( k), p1 ( k), Ts  .
 ) (84)
This function provides the variables of the process at the current sampling period. The filter is
implemented by the function Mouving_Average, having the input the switching voltage vector
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

33

components (vdopt,vqopt), length of the window of the filter L and the sample time period Ts, providing
the output of the moving window that corresponds to the filtered inputs (udftr, uqftr):

ud ftr , uq ftr  = Mouving _ Average[vd opt , vq opt , L, Ts ] . (85)

 

The numerical data calculated at each step Ts are stored in the DB named Data_Base_1, which
will continue to be used in DDC Algorithm presented in the next section.

3.4.2. DDC Algorithm

The DDC controls (ud, uq) are directly obtained from DB with MF (Figure 11) by the
interpolation techniques RBF – NN exposed in the seconnd section. The control low was generated
by means of a known set MF of operating points of the process (55) through an interpolation
technique AF (25). The resulting Algorithm 2 of the DDC strategy in pseudocode format is depicted
in Figure 19 that schematically is given by Figure 13. The main common programing syntax from
MPC and DDC has the same significance.
In first line there is loaded the data from database DB1, created with the MPC algorithm (Figure
18), where the support points of the control surfaces are described by:

u( a) = F(mm ( a), ( a)),

T
u = ud ftr ( a) uq ftr ( a)  ,
T (86)
F =  Fd Fq  ,
a = 1, , N obs .

As mention before, the control system is designed in two stages. In the the first, by MPC software
are simulated a set of motion trajectories. In database DB1 are stored in a set of grided operation
points of the process. Based on DB1 is simulated the DDC algorithm with which they were obtained
the results presented in the paper.
There is a particularity of the electric drive systems, because the controls udq(k) must be shaped
by the switching vector vdq of the power inverter (81). The set MF of operating points of the process
(55) is essential for the efficiency of the energy conversion system with power electronic converter
and electric machine. In that aim we selected from a large database a few set of the optimal energy
operation points defined by the speed, torque, votltge control of power converter, input and output
powers.
The parameters Pem (line 5) are used for the offline simulation of the process. The online
operation of the control algorithm is parameters free. These is the main advantage of the DDC
method.
In the main loop there are called the following functions:
• The PI_control which calculate the actual torque set point mref using the measured past speed ωmold
and the speed set point ωmref, and the tunning parameters Kp and Ki ;
• The matlab RBF_Interpolant function which interpolate the control surfaces (24) by scattered
support points contained in DB 1;
• The function Process_Model which calculate the actual process variables (id,iq,m,ω,) and the
powers p1,p2.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

34

Algorithm 2. The DDC main program

% Load Data_Base1
𝑓𝑡𝑟 𝑓𝑡𝑟
1. Load(‘DB1’,ωω’); Load(‘DB1’,mm’); Load(‘DB1’,𝑢𝑑 ’); Load(‘DB1’,’𝑢𝑞 );
% Initialization for the 8 possible values of the inverter voltages components
2. vd=[0;2;1;-1;2;-1;1;0]*UDC/3;
3. vq=[0;0;1;-1;0;-1;-1;0]* sqrt(3)/3* UDC; v=vd + vq*j;
% Initialization of the electromagnetic parameters
4. svs=[000; 100; 110; 010; 011; 001; 101; 111]; %Switching vector states inverter’s gate control
5. Pem=[mℓ, J, μ, Rs, Ld, Lq, ψPM]; %Declare process and machine parameters (*)
6. ω =ωref rated
;m max
=2.5*m rated
; % Initialization of the desired speed and maximum torque
7. ω old init
=ω ; id old
=0; iqold =0; udold-ref=0; uqold-ref=0; vdopt =0, vqopt =0 %Initializations
8. Ts=1e-4; Td=10s; N=Tdesired/Ts; Tint= 2 %Sampling period (Ts), desired simulation time (Tdesired), integration time (Tint)
9. Hm=8; dud=0; dud=0; g-opt=1e7; % Hm is shaping horizont, dud, duq, are the control corrections
10. Fd=RBF_Interpolant(ωω,mm,udftr) ; Fq=RBF_Interpolant(ωω,mm,uqftr) ; % Learning algorithm
11. for n=1:N
12. if (mod(n,Hm)) ==0) % the speed loop is the Hm time is slower then DDC algorithm
% Call function PI_control
13. [mref]=PI_control(ωref ,ωold ,mmax, Kp, Ki );
% Call interpolation algorithm by means of Data_Base 1 (speeds ωω, torques mm and controls udftr, uqftr).
14. ud= Fd (ωold,mref) ; uq= Fq (ωold,mref); % DDC law
15. end % end if
%Shaping the finite states of inverter
16. udref = ud-oldref+Tint(udref-vdopt); uqref = uq-oldref+Tint(uqref- vqopt);
17. ud-oldref= udref; uq-oldref= uqref;
18. for Hm=1:8
19. g=abs((udref+1* uqref)-v(ss));
20. if (g<g-opt)
21. g-opt=g;
22. ss_opt=ss;
23. end % end if
24. end % end for
25. vopt=v(ss_opt); vd opt= real(v(ss_opt)), vq opt= imag(v(ss_opt). % end of shaping
% Call Proces_Model
26. [id,iq,m,ω,p1,p2]= Proces_Model(vdopt,vqopt,idold,iqold,ωold ,Pem);
27. idold=id; iqold=iq; ωold=ω;
28. [udftr,uqftr]=Mouving_Average(vdopt,vqopt,L,Ts); % L is the length of the moving window filter
% Save process outputs in Data_Base_2
29. [idd(n),iqq(n), mm(n),ωω(n),uddftr,uqqftr,pp1(n),pp2(n)]=Data_Base_2=(id,iq,m,ω,udftr,uqftr,p1,p2);
30. end % end for

Figure 19. The DDOC control algorithm of a PMSM drive control.

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

35

The control u(a)= ud(a)+juq(a) are the query of the actual points (ωmold, mref) for the surfaces Fd and
Fq . These controls are sampled from the continuous time domain. The inverter is controlled by the
finite states switching functions svs . The transformation of the controls ud and uq from continuous
time domain in the finite states domain of the inverter’s switching vectors vd and vq is a synthesis
process.
The actual control u(a) at the step a is synthetized by a finite set of the switching vectors v=vd+jvq
as follows

u( a) = {vk opt (r ), vk opt (r + 1), vk opt (r + H m )},

r = 1,..., H m , . (87)
a = 1,..., N obs .

where Hm is the horizon of modulation and each vopt(k) vectors are determinates with a searching
procedure, minimizing the g criteria function. The dimension of the Hm horizon was chosen by
analogy with the classical modulation.

4. Illustrative Case Study of DDC of a PMSM Drive System

An industrial PMSM from a drive system supplied by a power inverter is used to illustrate the
proposed DDC approach by comparison with the MPC control in the order to check the optimal
energy operation. The PMSM rated data are depicted in Table 2.

Table 2. PMSM rated specifications data.

Symbol Description Values

PN [W] Rated power 5500
UN [V] Rated voltage 325
mN [Nm] Rated torque 35
IN [A] Rated current 10.6
ωmN [rad/sec] Rated speed 157
Rs [Ω] Rotor resistence 0.65
Ld [H] Direct axis inductance 0.0082
Lq [H] Quadrature axis inductance 0.0082
J[kg∙m2] Total inertia of the PMSM drive 0.5
zp Stator pole pairs 2

The illustrative case study is done in Matlab-Simulink software in command-line line

implementation programs. For simulation it has been adopted the sampling time period Ts=10-4s.
Also, the speed loop is designed by choosing a linear control law implemented by PI controller
designed by pole placement method, resulting the parameters Kpω=30 and Kiω=0.1. As aprioric data
KDB there are considerate the PMSM parameters (Table 2), rated data for a conventional power
invertor with DC link voltage UDC=520V, the PMSM operate in a closed room.
A first objective is the designing of database used for DDC steategy based on the high efficiency
data colected from the MPC stratefy running of PMSM. The second aim is to check the results of a
comparative analysis obtained by MPC and DCC in rated conditions. To study a high impact of the
DDC method, a third scope is to made a comparative analysis of the reults obtained via MPC and
DCC strategies in mismatch conditions. Finally, a performance comparative analysis of the privious
results obtained via MPC and DDC strategies for rated and mismach conditions is done in the order
to demonstrate that the developed DDC method in the paper is a relevant topic for future research.

4.1. Database Learning Design

Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

36

The essential point in DDC method is the DB design. Firstly, there are the bootstrap DB obtained
from process design theoretical data and parametric offline of the process.
From this large knowledge and data DB we extracted the control (working) DB. The control DB
containes a few optimal efficiency steady-state opearation points of the process (around 250),
obtained in exagenous conditions. In our case the exagenous conditions are the ambiental
temperature T0 and rated flux ψPM. Tacking into account the ambiental conditios the control becomes:

u = AF (m , m, T o , PM ) . (88)
In consequence, the control DB must be extracted, and the RBF interpolation block will have four
inputs and two outputs controls.
In this paper we applied the direct searching corrections of the controls in the sens of maximum
energy efficiency.This corections are global or local.
A gridded experimental data points was obtained by simulating the MPC control algorithm,
providing an admissible set of the operation points of the process selected in steady-state regime for
the highest efficiency values.
There were adopted voltage global correction on the entire DB:

udDDC = udMPC kd
 DDC , (89)
uq = uq kq
MPC

where the factors kd=1.04 and kq=1 are selected in the order to improve the DDC performances.
The local DB corection are applied around a steady-state points. This kinds of corrections are
applied mainly online in steady-state operation, in the sense of searching the maximum energy
efficiency, in the actual exogenous conditions,
Based on this set experimental data, in Figure 20 it was built the discrete implicit surfaces of the
high efficiency DDC control law  DDC = f (mDDC , mDDC ) , which contains around 250 points of
steady-state higher efficiency operation given by MPC strategy.

Figure 20. The efficiency surface vs. speed and torque.

A more attention must be paid for representing the control sufaces in Figure 21 and Figure 22 of
the control implicit surfaces ud / q = f (mDDC , m DDC ) .
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

37

Figure 21. The d-axis voltage surface vs. speed and torque.

Figure 22. The q-axis voltage surface vs. speed and torque.

4.2. Study on Rated Data Conditions

In this section is provided a compartive analysis of the results obtained via MPC and DDC in
rated conditions provided by the model parameters of PMSM from Table 2.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

38

Adopting a speed step reference of mref = 175rad / s , in Figure 23 is depicted the speed tracking
results obtained by MPC and DDC strategies. The speed provided by DDC method mDDC ad MPC
one mMPC present close responses, but the DDC has a sensible lower transient regime.

Figure 23. The speed tracking results obtained for MPC and DDC control strategies.

The current results provided by MPC and DDC strategies are illustrated in Figure 24. In the case
of d-axis current component, the MPC law offers a small value then the corresponding one of DDC
law in absolute values id
MPC
 idDDC . After a different dynamic regime, the q-axis current component
presents close values in steady state regime for for both DDC and MPC controllers.

Figure 24. The current results obtained for MPC and DDC control strategies.

On the shaft of the PMSM dtive is applied a step load torque m = 17.5Nm . The load and
electromagnetic torques provided by MPC/DDC laws are represented in Figure 25. Excepting the
startup process, in dynamic regime the electromagnetic torque obtained by DDC si smaller then one
given by MPC m
DDC
 m MPC . However, the speed transient leads to have a higher dynamic torque
regime for DDC method.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

39

Figure 25. Load and electromagnetic torques for MPC and DDC control strategies.

Figure 26 indicates the voltages obtained by both MPC and DDC strategies udMPC
/q
and udDDC
/q
. As
a consequnce of the increasing of the q-axis voltage component provided by DB, this difference is
seened in the figure, having udMPC  udDDC . For d-axis, the votage component have closed values for
MPC and DDC strategies.

Figure 26. Voltage components for MPC and DDC control strategies.

To have a concrete appreciation of the current tracking or voltage delivered by the control
system of PMSM, in the following there are evaluated the current/voltage magnitude computed
according to:

I = I d2 + I q2 , (90)
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

40

U = ud2 + uq2 , (91)

which are graphically illustrated in Figure 27 and Figure 28. It can be seen that, after the
startup, during the transient regime the DDC method provide a lower value then de
corresponding one given by MPC strategy I
DDC
 I MPC , and in steady-state regime
becomes grater I
DDC
 I MPC .

Figure 27. Current magnitude for MPC and DDC control strategies.

Figure 28. Voltage magnitude for MPC and DDC control strategies.

The efficiency of PMSM obtained by MPC and DDC control strategies in depicted in Figure 29.
As expected, the new control technology provides advanced benefits as parameter free and high
efficiency objective. Thus, the dynamic efficiency compared with the classical one steady state
efficiency is greater    MPC .
DDC
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

41

Figure 29. Efficiency of PMSM for MPC and DDC control strategies.

To evaluate the efficiency of the PMSM on the enitre simulation domain, there is used the
average efficiency defined according to:
N obs

  ( a) (92)
 = av a =1
obs ,
N
which is graphically represented in Figure 30.

Figure 30. Average efficiency of PMSM for MPC and DDC control strategies.

Finally, it is important to ases the phase currents. Since the PMSM has a symmetric mecahical
and electrical structure, the system of currents is also symetrical, which is why only the current on
the first phase iaMPC / DDC will be represented as in Figure 31. The peak of the phase current proided
by DDC method has a high value then one given by MPC law.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

42

Figure 31. Phase current of PMSM for MPC and DDC control strategies.

A first conclusion is that the DDC strategy may operate without any mathematical model, being
also parameter-free. The key of DDC strategy operation is related on database design, learning
process and available computational resources. A more important appreciation is that DDC improve
both tracking refecences and efficiency. However, the DDC strategy remains stable in the imposed
simulation scenario.

3.3. Study on Mismatch Conditions

Often, in practice the dq PMSM model with constant lumped parameters is just an idealization
which cannot always be fulfilled.
In the following is considered there the next variations of the parameters of the dq model of
PMSM due to demagnetization and temperature increasing, respectively, as follows:

 PM → 0.8 PM
 , (93)
 Rs → 1.4 Rs

otherwise, all the conditions remain unchanged.

In this case, the speed tracking results obtained by MPC and DDC strategies ate illustrated in
Figure 32. Now, again, the speed obtained by DDC strategy mDDC has a shorter dynamic regime
then the one given by MPC law mMPC . The speed response is even better than the one obtained in the
rated scenario. For MPC strategy, the results get worse once with parameters mismatch.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

43

Figure 32. The tracking results of speed obtained via MPC and DDC control laws.

For current results given by MPC and DDC laws, both transient and steady-state regimes are
different as is depicted in Figure 33. The d-axis current component delivered by MPC controller and
the one of DDC law in have different values in dynamic regime, and closed in steady-state regime.
On q-axis, the current given by MPC and DDC have different values in both dynamic and steady-
state regime.

Figure 33. The dq current obtained for MPC and DDC control laws.

Applying the same step load torque m = 17.5Nm , the load and electromagnetic torques
obtained by MPC/DDC strategies have different dynamics as is graphically depicted in Figure 34.
The dynamic torque given by DDC has a higher value then the one provided by MPC m DDC  m MPC
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

44

, but acts in a short transient. Thus, the evolution of electromagnetic torques is done with different
dynamics until the main disturbance represented of the load torque is rejected.

Figure 34. Load torque disturbance rejection for MPC and DDC control strategies.

The voltages provided by both MPC and DDC strategies udMPC

/q
and udDDC
/q
are shown in Figure
35. On d-axis, the voltage components present appropiate values, with small differences in dynamic
regime. Important differences are founded on q-axis, where udDDC  uqMPC has a significant value,
mainly in steady-state regime.

Figure 35. Voltage obtained by MPC and DDC control strategies.

The magnitude of current and voltage are represented in Figure 36 and Figure 37, respectively.
The current magnitude obtained by DDC law is allways smaller then the one given by MPC algorithm
I DDC  I MPC . Mostly, a significant increasing of the voltage magnitude for DDC algorithm then MPC
controller is obtained U DDC  U MPC .
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

45

Figure 36. Current magnitude given by MPC and DDC control algorithms.

Figure 37. Voltage magnitude given by MPC and DDC control algorithms.

The efficiency of PMSM obtained by DDC law is substantially increased the the one provided
by and DDC strategy    MPC in both dynamic and steady state regimes, as illustrated in Figure
DDC

38. This is a consequnce of the fact that DDC method is designed by using the high-efficincy data
colected from MPC operation. This is a major objective done in this robust scenario.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

46

Figure 38. Efficiency of PMSM obtained by using MPC and DDC control algorithms.

As expected the average efficiency for the entire simulation time is increased for DDC method
 av − DDC
  av − MPC as is graphically depicted in Figure 39.

Figure 39. Average efficiency of PMSM provided by using MPC and DDC control algorithmes.

In this scenario, the phase current obtained by DDC strategy has a smaller peak value then the
one delivered by MPC algorithm in both transient and steady-state regimes as is shown in Figure 40.
However, a small peak value will generate reduced power losses, and then the increasing of the
efficiency of PMSM. The DDC strategy is designed by using a database learning that no not contain
overcurrents, feature which will maintain the phase current into a minimal range.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

47

Figure 40. Phase current for MPC and DDC control algotihms.

The study developed in this section shows the results given by DDC strategy under robustness
conditions of PMSM parameters mismatch. The main objective of the efficiency increasing has been
successfully shown. However, this objective leads secondary to decrease the dynamic regime of
speed tracking. Compared with MPC strategy, the DDC law requires a higher voltages
magnitude/components which must take into account the voltage limit levels of the PMSM drive.
Also, the DDC strategy leads to obtain a smaller peak value of the phase current that will generate
lower level of power losses.

3.4. Comparative analysis of MPC and DDC results

The proposed DDC method proves to have a stable operation on the rated an mismatch cases
and, also offers several benefits. For a relevant comparison of the above discussed cases, there are
used the next criteria:
- Settling time of the mechanical speed – tst;
- Average efficiency on the entire simulation time– ηav;
- Peak value of the a phase current in steady-state regime– Iapk;
- THD of the a phase current computed in steady-state regime, for ten harmonics, including the
fundamental of the frequency 55.66 Hz, for a comman time domain pf MPC and DDC strategies, 4-5
s time.
The comparative results based on the above criteria of the MPC and DDC strategies are
summarized in Table 3.

Table 3. The comparative evaluation of MPC and DDC strategies.

Test Control tst [s] ηav Iapk [A] THDa[%]

strategy
Rated conditions MPC 2.52 0.730 6.55 0.06
DDC 2.50 0.742 6.96 0.17
Mismach MPC 3.68 0.612 7.86 0.08
conditions DDC 2.28 0.722 6.44 0.13
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

48

It is obvious that in both tests, compared with MPC law, the DDC strategy leads to decreasing
of the settling time of the mechanical speed t st − DDC  t st − MPC . In rated conditions there is is obatined
a small diffrence, but in the mismatch test the DDC method leads to a setting time with 61% lower
then the one obtained by MPC strategy, which is a high advantage. For rated conditions, the
increasing of the average efficiency is small for DDC strategy  av − MPC   av − DDC , being around 1.2 %
from the one resulted in rated conditions with MPC controller. A high increasing of the DDC average
efficiency  av − DDC   av − MPC , about 11 % results for mismatch test, which is the major successfully,
achieved objective. In addition, the peak value of the phase current for rated conditions is smaller for
DDC, I apk − DDC  I apk − MPC , and higher in the mismatch case I apk − DDC  I apk − MPC . As a consequnce of
interpolation and digital signal processing, the THD of the phase current obtained by DDC law
remais low for both rated/mismach cases studied.
We can conclude that DDC strategy persevere better the high efficiency objective in the both
rated and robust scenarios then the classical MPC law. Nevertheless, the settling of speed responses
has been also improved in the both tests.
Finally, it should be noticed the results of the performances presented in Table 3 are open to be
improved since the results depends on the learning process and on software and hardware available
resources.

5. Conclusions
The many faces of the macroscopic energy conversion systems may be characterized by a set of
generalized variables as effort and flow. This approach allows a physical – based mathematical
equation modeling developed in the order to increase the efficiency of energy conversion.
In spite of the high performances of actual devices and actuators for the energy transformation,
there are enough resources for the improvements of their efficiency, especially in the operation points
at low effort and high flow or respectively at high effort and high flow.
A transition from MBC to MFC is required for ensuring the non-sensitive control approach.
The idea of the paper is to build the surfaces of the DDC law in terms of effort and flow. Starting
from a large knowledge DB we built a control DB containing a reasonable optimal efficiency steady
state operation points of the process. A RBF interpolative system has been trained in order to extract
from this surfaces the controls demanded for any operation points. It results that the control efficiency
is largely dependent from the consistency of the colected data.
The AC drives is a field that emphasis the largely developed fast process in industry and
residence. The performances of this DDC method were evaluated based on a PMSM drive system
simulated by MPC methodology. The DB was built in Matlab simulation environment using a specific
algorithm for the MPC control of PMSM drive and the algorithm of DDC. The comparison criteria
were the efficiency of the energy process conversion (the ratio of the output and input powers), the
quality of the control for speed and effort and the current THD. Due to the PMSM windings resistance
variations and to the degree of the magnetic circuit depreciations during the time, the DDC efficiency
control is better than MPC control, because the MPC is model parameters dependent.
In many research papers in the field has been demonstrated that the direct DDC strategy is
superior in reduction the bias whereas the indirect MPC one gives beter variance. The results shown
in Figure 36 and Figure 38 confirm the fact that the MPC control gives better variance but the DDC
gives beter efficiency and bias to process parameters modifications.
Our DDC technique, being process parameters free is more general then classical process model
control as MPC, in the sense that does not depend on the nature of process (electromagnetic , thermic
or mechanic). If we reefer on the electric driven systems, DDC technic is the same for all type of
motors (asynchronous, synchronous or others).
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

49

The DDC method with appropriately designed management make possible to realize a high
level of the energy transformation quality that is fast dynamic response, high energy efficiency and
low parametric sensitivity.
The muliple aspects disscused and successfully solved in the paper shows that the novel
tehnology developed in the paper is still an open problem which can provides realistic improved
results in a future promissing approach.

Author Contributions: Conceptualization, M.C. and I.B.; methodology, I.B. (Ion Bivol); software, M.C. (Madalin
Costin); validation, M.C. and I.B.; formal analysis, I.B.; investigation, I.B.; resources, M.C.; writing—original draft
preparation, M.C.; writing—review and editing, I.B.; visualization, M.C.; supervision, I.B. All authors have read
and agreed to the published version of the manuscript.
Funding: This research was partially supported under Contract no. RF 2462/31.05.2024 “Performance Evaluation
of a Novel Control Strategy for Electromagnetic Energy Conversion Processes based on RBF‐NN techniques, by
comparison with Established MPC‐type Methods, with respect to PMSM Motors” from “Supporting and
Developing Research, Innovation and Technology Transfer activities at the University “Dunărea de Jos” of
Galați”.
Data Availability Statement: Data availability is not applicable to this article as the study did not report any
data.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations
AC Alternative Current
AIMs Artificial Intelligence Methods
DB DataBase
DDC Data Driven Control
DC Direct-current
EDSs Electric Drive Systems
EMF ElectroMotive Force
HVAC Hydro, Ventilation, Air Conditioning
MBC Model Based Control
MFC Model Free Control
MPC Model Predictive Control
NN Neural Network
PI Proportional-Integral controller
LPF Lower Pass Filter
PMSM Permanent Magnet Synchronous Machine
RBF Radial Basis Function
RL Reinforcement Learning
THD Total Harmonic Distorsion
QP Quadratic Programming
WB White-Box

Nomenclature
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

50

(a,b,c) Three phase coordinates

AF Set of the weights of the RBF-NN
(α,β) Fixed coordinates
αm Electric position of the rotor
dMBC/MPC Set of processing data obtained by MBC/MPC
doMBC/MPC Steady-state set of processing data obtained by MBC/MPC
doe Set of data obtained by grid learning
Cc,Cd Continous/discrete cost
DF DB of knowledge and data of the process
Δ Unit-delay
e Effort
ed/q d/q electrical EMF components
E f , RMS , En , RMS RMS value of fundamental/ Harmonics components of the effort e
EN Set of envirovment data
E1,E2 Input/output energy
emf Motion ElectroMotive Force (EMF)
ε Mean square error
η,ηav Efficiency/Average efficiency
ξ Discrete noise of the process state model
F Operator of control law
f,g Nonlinear function of the state representation
Fi Explicit control law
fε Implicit represenation of the process model
fu Useful force
f1/2 Components of surface grid representation
fm/p Multi-valued map of torque/output power
f ‘m/p Single-valued map of torque/output power
ψPM Permanent magnet flux
φi Gaussian distribution
ΦMBC/MPC Training data set given by MBC/MPC
g Cost function
Gi Implicit control law
g1,g2 Generalized variables of trajectories
Γm/p Set of speed and torque/output power and efficiency
Γ ’m/p Set of speed and torque/output power and maximum efficiency
h/h Output of the space state model
Hm Horizon of command shaping searching
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

51

H/Hmax Set of efficiency/maximum efficiency

HN Set of design data
i Index that corresponds to the combination of switching
(id,iq) Direct and quadrature current components
i Current phasor
Iapk Peak value of the a phase current
I Input of the RBF-NN
J Motal inertia
kd/q Corection factor of voltage
ke Back motion EMF constant
KDB Knowledge data
l Index of the number of hidden neurons of the RBF-NN
(Ld,Lq) Direct and quadrature inductances
Lw Filter length
LPF Lower Pass Filter
λ Generalized variables for transformation of PMSM model
Λm/p Set of the cartezian product of speed and torque/output power
m,mℓ Electromagnetic/load torque
M Set of electromagnetic torque
MF Set of training data for RBF-NN

K db Set of the aprioric knowledge

μ Center points of Gaussian function
μf Friction factor of the load
N Number of trajectory points
Nω/m/p/h Number of points of the set of speed,torque,power,efficiency, respectively
Nobs Number of observation points
Q Flow rate
θ Vector of parameters of the process state model
P Pressure
P Set of output power
p1,p2 Input/output power
e
Set of electrical parameters
m
Set of mechanical parameters
em
Set of electromechanical parameters
RN Set of rated data
Rs Stator resistance
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

52

RMS Root mean square

Re{},Im{} Real/imaginary part of a given qunatity
s Flow
Sabc Set of the switching states
σ Basis width vector of the Gaussian distribution
t Time
t0,t1 Start/end time
Td Desired final time
Ts Sampling time period
0
T Temperature od stator winding of PMSM
−1
T , T Direct/Indirect matrix of the Clarke transformation
Tdq , Tdq−1 Direct/Indirect matrix of the Park transformation
Ti Power inverter matrix
u Control input
û Any other control then the optimal one
u Control obtained via grid surface interpolation tehnique
uabc Three-phase voltages
udq dq voltages
uftr Filter input
uRBF Output of the RBF-NN
u Voltage phasor
UDC DC bus voltage
(ud,uq) Direct and quadrature voltage components
(ud ftr/nftr
, Filtred/non-filtred commande of the power inverter
ftr/nftr
uq )
v Linear velocity
vd/q Vectors of d/q components of power inverter
vcdq/ opt − DDC Command/optimal voltage of the DDC algorithm

ω e, ω m Electrical/mechanical angular speed

ωm,b Based mechanical speed
ωm,max Maximal mechanical speed
ωmN Rated mechanical speed
x State of the state model
y Output of the state model
y Output of the interpolation method
yftr Filter output
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

53

z Generic variable used for the appling of the forward Euler forward
method
zp Stator magnetic poles pairs
w Weights
W Database operator of data processing
Ω Mecahical speed set

References
Sul, S.K. Control of Electric Machine Drive Systems, John Wiley and Sons: Hoboken, NJ, USA, 2011.
1. Vas, P. Sensorless Vector and Direct Torque Control, Oxford University Press, NY, USA, 1998.
2. Rodriguez, J.; Cortes, P. Predictive Control of Power Converters and Electrical Drives, John Wiley & Sons, Ltd.:
West Sussex, UK, 2012.
3. Holtz, J. Advanced PWM and predictive control—An overview. IEEE Trans. Ind. Electron. 2016, 63, 3837–
3844.
4. Casadei, D.; Serra, G.; Tani, K. Implementation of a direct control algorithm for induction motors based on
discrete space vector modulation. IEEE Trans. Power Electron. 2000, 15, 769–777.
5. Vafaie, M.; Dehkordi, B.; Moallem, P.; Kiyoumarsi, A. Minimizing torque and flux ripples and improving
dynamic response of PMSM using a voltage vector with optimal parameters. IEEE Trans. Ind.
Electron. 2015, 63, 3876–3888.
6. Vazquez, S.; Leon, J. I.; Franquelo, L.G.; Rodriguez, J.; Young, H. A.; Marquez, A.; Zanchetta, P. Predictive
control: a review of its applications in power electronics. IEEE Ind. Electron. Mag. 2014, 8, 16-31.
7. Vazquez, S.; Rodriguez, J.; Rivera, M.; Franquelo, L.G.; Norambuena, M. Model predictive control for
power converters and drives: Advances and trends. IEEE Trans. Ind. Electron. 2017, 64, 935–947.
8. Rodríguez, J.; Kennel, R.M.; Espinoza, J.R.; Trincado, M.; Silva, C.A.; Rojas, C.A. High-performance control
strategies for electrical drives: An experimental assessment. IEEE Trans. Ind. Electron. 2011, 59, 812–820.
9. Zhang, K.; Fan, M.; Yang, Y.; Zhu, Z.; Garcia, C.; Rodriguez, J. Field enhancing model predictive direct
torque control of permanent magnet synchronous machine. IEEE Trans. Energy Convers. 2021, 36, 2924–
2933.
10. Schwenzer, M.; Ay, M.; Bergs, T.; Abel, D. Review on model predictive control: an engineering
perspective. Int J Adv Manuf Technol. 2021, 117, 1327–1349.
11. Richter, J.; Doppelbauer, M. Predictive trajectory control of permanent-magnet synchronous machines with
nonlinear magnetic. IEEE Trans. Ind. Electron. 2016, 63, 3915-3924.
12. Djerioui, A.; Houari, A.; Machmoum, M.; Ghanes, M. Grey wolf optimizer-based predictive torque control
for electric buses applications. Energies 2020, 13, 5013.
13. Siami, M.; Khaburi, D.A.; Abbaszadeh, A.; Rodriguez, J. Robustness improvement of predictive current
control using prediction error correction for permanent-magnet synchronous machines. IEEE Trans. Ind.
Electron. 2016, 63, 3458–3466.
14. Hu, F.; Luo, D.; Luo, C.; Long, Z.; Wu, G. Cascaded robust fault-tolerant predictive control for PMSM
drives. Energies 2018, 11, 3087.
15. Singh, B.; Jain, P.; Mittal, A.P.; Gupta, J.R.P. Torque ripples minimization of DTC IPMSM drive for the EV
propulsion system using a neural network. J. Power Electron. 2008, 8, 23–34.
16. Wang, W.; Liu, Y.; Chen, H.; Gao, J.; Zhu, S.; Zhou, R. Improved Rotor Flux-Based SMO and RBF-PID
Control Strategy for PMSM. Actuators 2023, 12, 327.
17. Qian, W.; Panda, S.K.; Xu, J.X. Torque ripple minimization in PM synchronous motors using iterative
learning control. IEEE Trans. Power Electron. 2004, 19, 272–279.
18. Wang, G.; Xie, J.; Wang, S. Application of Artificial Intelligence in Power System Monitoring and Fault
Diagnosis. Energies 2023, 16, 5477.
19. Schenke, M.; Kirchgassner, W.; Wallscheid, O. Controller design for electrical drives by deep reinforcement
learning: A proof of concept. IEEE Trans. Ind. Informat. 2020, 16, 4650– 4658.
20. Xie, H.; Wang, F.; Xun, Q.; He, Y.; Rodriguez, J.; Kennel, R. A low-complexity gradient descent solution
with backtracking iteration approach for finite control set predictive current control. IEEE Trans. Ind.
Electron. 2020, 69, 4522-4533.
Prag, K.; Woolway, M.; Celik, T. Toward data-driven optimal control: a systematic review of the landscape. IEEE Access
2022, 10, 32190-32212.
21. Hou, Z. S.; Wang, Z. From model-based control to data-driven control: survey, classification and
perspective. Inf. Sci. 2013, 235, 3–35.
Preprints.org (www.preprints.org) | NOT PEER-REVIEWED | Posted: 7 October 2024 doi:10.20944/preprints202410.0516.v1

54

22. Markovsky, I.; Dörfler, F. Behavioral systems theory in data-driven analysis, signal processing, and control.
Annu. Rev. Control 2021, 52, 42-64.
23. Markovsky, I.; Dörfler, F. Data-driven dynamic interpolation and approximation. Automatica 2022, 135,
110008.
24. Hao, W. Data‐Driven Control with Learned Dynamics. Ph.D. Dissertation, Clemson University, South
Carolina, USA, August 2020.
25. Dörfler, F.; Coulson J.; Markovsky, I. Bridging direct and indirect data-driven control formulations via
regularizations and relaxations. IEEE Trans. Autom. Control 2023, 68, 2023.
26. De Persis, C.; Tesi, P. Formulas for data-driven control: stabilization, optimality, and robustness. IEEE
Trans. Autom. Control 2020, 65, 909–924.
27. Kadali, R.; Huang, B.; Rossiter, A. A data driven subspace approach to predictive controller design. Control
Eng. Pract. 2003, 11, 261–278.
28. Coulson, J.; Lygeros, J.; Dorfler, F. Distributionally robust chance constrained data-enabled predictive
control. IEEE Trans. Autom. Control 2021, 67, 3289-3304.
29. Abolpour, R.; Khayatian, A.; Dehghani, M. Simultaneous model prediction and data-driven control with
relaxed assumption on the model. ISA Transactions 2024, 145, 225-238.
30. Yang, L.; Ma, A.; Li, D.; Xi, Y. Input-mapping based data-driven model predictive control for unknown
linear systems with bounded disturbances. Automatica 2023, 153, 111056.
31. Yang, L.; Li, D.; Ma, A.; Xi, Y.; Pu, Y.; Tan, Y. Input‐mapping based data‐driven model predictive control
for unknown linear systems via online learning. International Journal of Robust and Nonlinear Control 2022.
32. Coulson, J.; Lygeros, J.; Dorfler, F. Data-Enabled Predictive Control: In the Shallows of the DeePC. In
Proceedings of the 18th European Control Conference (ECC), Naples, Italy, 25-28 June 2019, pp. 307-312.
33. Brosch, A.; Hanke, S.; Wallscheid, O.; Bocker, J. Data-driven recursive least squares estimation for model
predictive current control of permanent magnet synchronous motors. IEEE Trans. Power Electron. 2021, 36,
2179–2190, 2021.
34. Carlet, P. G.; Favato, A.; Bolognani, S.; Dörfler, F. Data-driven continuous-set predictive current control for
synchronous motor drives. IEEE Trans. Power Electron. 2022, 37, 6637-6646.
35. Carlet, P. G.; Favato, A.; Torchio, R.; Toso, F.; Bolognani, S.; Dörfler, F. Real-time feasibility of data-driven
predictive control for synchronous motor drives. IEEE Trans. Power Electron. 2023, 38, 1672-1682.
36. Xie, Y.; Tang, X.; Song, B.; Zhou, X.; Guo, Y. Data-driven adaptive fractional order PI control for PMSM
servo system with measurement noise and data dropouts. ISA Transactions 2018, 75, 172-188.
37. Guo, Q.; Zhang, C.; Li, L.; Zhang, J.; Wang, M. Maximum Efficiency per Torque Control of Permanent-
Magnet Synchronous Machines. Appl. Sci. 2016, 6, 425.
38. Cavallaro, C.; Tommaso, A.O.; Miceli, R.; Raciti, A.; Galluzzo, G.R.; Trapanese, M. Efficiency Enhancement
of Permanent-Magnet Synchronous Motor Drives by Online Loss Minimization Approaches. IEEE Trans.
Ind. Electron. 2005, 52, 1153–1160.
39. Ni, R.G.; Xu, D.G.; Wang, G.L.; Ding, L.; Zhang, G.Q.; Qu, L.Z. Maximum Efficiency Per Ampere Control
of Permanent-Magnet Synchronous Machines. IEEE Trans. Ind. Electron. 2015, 62, 2135–2143.Ronggang Ni,
Dianguo Xu, Gaolin Wang, Li Ding, Guoqiang Zhang, Lizhi Qu. Maximum efficiency per ampere control
of permanent-magnet synchronous machines. IEEE Trans. Ind. 2015, 62, 2135–2143.
40. McCarthy, J.; Minsky, M.L.; Rochester, N.; Shannon, C.E. A Proposal for the Dartmouth Summer Research
Project on Artificial Intelligence, August 31, 1955. AI Magazine. 27, 4 (Dec. 2006), 12.
41. Liu, J. Radial Basis Function (RBF) neural network control for mechanical systems: design, analysis and Matlab
simulation. Springer Science & Business Media, 2013.
42. Kaminski, M. Nature-Inspired Algorithm Implemented for Stable Radial Basis Function Neural Controller
of Electric Drive with Induction Motor. Energies 2020, 13, 6541.
43. Astrom, K.J.; Wittenmark, B. Computer-Controlled Systems: Theory and Design, 3rd ed. Prentice Hall: Upper
Saddle River, NJ, USA, 1997.

BoDisclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of
the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim
responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred
to in the content.