HUNGARIAN JOURNAL OF

INDUSTRY AND CHEMISTRY

Vol. 44(2) pp. 105–111 (2016)
[Link]
DOI: 10.1515/hjic-2016-0013

MODELLING A THREE-PHASE CURRENT SOURCE INVERTER

L ÁSZLÓ N EUKIRCHNER * AND A TTILA M AGYAR

Department of Information Technology, University of Pannonia, Egyetem u. 10, Veszprém, 8200,

HUNGARY

A current source inverter model has been developed in the given paper that is constructed from six LTI models
for the different switching modes. The overall model is in a piecewise affine form that supports the use of model
predictive control. The model has been verified against engineering expectations and its open-loop performance
shows that it is a promising basis of model predictive control structures.

Keywords: model predictive control, current source inverter, space vector modulation, modelling,
simulation

1. Introduction current harmonic distortion, and could be

straightforwardly implemented in digital systems [1]. In
the second section, the structure of the investigated
In industrial applications power electronic converters inverter is presented. In the next section, a detailed
are divided into two major classes, namely voltage open-loop model is presented, and in the fourth section
source (VSI) and current source inverters (CSI). As a a predictive strategy is chosen for controlling SVM.
VSI employs modulating the voltage wave through the
inverter’s legs, its counterpart relies on modulated
current waveforms. To operate a VSI the inverter’s 2. Inverter Structure
switching devices should operate in a specified
alternating modulation to avoid shorting the inverter’s
The current source inverter normally requires three-
legs, while its current source counterpart could embed
phase inductive and capacitive filtering (Lf and Cf,
this operation into the modulation. Thyristor rectifiers
respectively), improving the load current and voltage
suffer from several drawbacks, e.g. high-input current
waveform’s harmonic distortion, and in thermistor
harmonic distortion or a low power factor. Pulse width
appliances assisting the commutation of the switching
modulated current-source inverters, on the other hand,
devices. The legs of the device should include
are free from the above-mentioned drawbacks. In such
protective diodes, otherwise the absence of a current
cases, a smoothing LC filter is placed on the AC side
path introduces a high-voltage spike, causing damage to
and a choke inductor DC side to reduce the harmonic
the switching devices [2]:
injection resulting from the current modulating PWM
operation. In the case of sinusoidal supply voltages, an • The switching devices used are symmetrical,
input power factor with unit value can be achieved if they do not require freewheeling diodes.
both requirements are met. In conventional control • The current source inverter produces a three-
structures, PWM-CSR is usually operated by off-line phase pulse-width modulated current instead
patterns and look-up tables like in trapezoidal of a pulse-width modulated voltage. As such,
modulation and selective harmonic elimination. The use a capacitor needs to be installed on the output
of these techniques simplifies the calculation of the gate for filtering purposes. The high voltage
control signals with the special requirements of applied derivative is associated with the VSI.
switches like IGBTs. Input filtering, on the other hand, • The rate at which the DC current rises is
may result in oscillatory transient behaviour. This can limited by the DC choke, allowing sufficient
be attenuated by adding damping resistive elements, but time for the protection circuit to function in
using them also reduces the overall efficiency. Space the case of a short circuit at the inverter output
vector modulation (SVM) has proven to result in a high terminals.
voltage gain, reduced switching frequency and low • As a restraint, the DC link current cannot be
instantaneously changed, which resulted in a
reduction in dynamic performance.
*Correspondence: [Link]@[Link]

Unauthenticated
Download Date | 7/5/17 4:28 AM
106 NEUKIRCHNER AND MAGYAR

Figure 1. Current source inverter with an LC filter and RL load.

Table 1. Switching states of SVPWM modulated CSI.

3. Open-Loop Switched Model with Space
Vector Modulation active
iA iB iC
Space
switches vector
S1, S4
Zero
3.1. Space Vector Formulation States
S2, S5 0 0 0 I0
S3, S6
The switched model of the three-phase CSI supplying a S6, S1 iDC -iDC 0 I1
three-phase load through an LC filter and RL load is S1, S2 iDC 0 -iDC I2
Active S2, S3 0 iDC -iDC I3
shown in Fig.1. The LC filter is a low-pass filter
States S3, S4 -iDC iDC 0 I4
included with the three-phase CSI to decrease the THD S4, S5 -iDC 0 iDC I5
by eliminating higher order harmonics. The S5, S6 0 -iDC iDC I6
implemented LC filter includes a resistor Rf in series
with a capacitor Cf. This resistor provides a passive These switching states can be classified as zero
damping solution and it was recommended to be within switching states (I0) and active switching states (I1, ...,
the range of 0.5 to 1 Ω. Moreover, the resistor RDC I6). In addition, there are three zero states which implies
represents any start-up pre-charging resistance or any vertical conduction of the inverter leg. This implies that
resistance in series with the DC voltage source VDC. the source is shorted, leading to iA = iB = i C = 0. This
For constructing the model the following assumptions operating mode is often referred to as the bypass
are made: operation. Assuming that the operation of the inverter is
• The LC filter and the RL load’s circuit three-phase balanced then Eq.(1) applies
components are symmetric.
• Only a continuous conduction mode is iA + iB + iC = 0 (1)
assumed.
• Each time, the switch state index s is known. According to Fig.2, the five active vectors can be
• The switching devices used for modulation are expressed as:
assumed to be ideal, with zero losses and
switching is considered to be instant. ! !
! ! !!! ! ! !
𝑰! = 𝐼!" e , k = 1, 2, …, 6 (2)
• No external disturbance is assumed. !
• No magnetic coupling between the
inductances is assumed. The active and zero vectors are assumed to be
With the above assumptions, the equation system stationary, on the contrary, the current reference vector
of the inverter, filter and load for all the switching states Iref in Fig.2 rotates with an angular velocity ω = 2πf, and
!
can be created, as different state-space models [3, 4]. θ 𝑡 = ! ω 𝑡 𝑑𝑡 + θ 0 , where f is the fundamental
Under the above-mentioned constraints, the three-phase frequency of the inverter output current and θ is the
current source inverter has nine switching states as can angular displacement. A desired vectorial length and
be seen in Table 1. angle, Iref can be calculated from the neighbouring space

Hungarian Journal of Industry and Chemistry

Unauthenticated
Download Date | 7/5/17 4:28 AM
MODELLING A THREE-PHASE CURRENT SOURCE INVERTER 107

Figure 3. Calculation of Iref in Sector I, where the

Cartesian coordinate system represents the imaginary
and real axes.
Figure 2. A typical space vector diagram for the CSI.
I1, ..., I6 are the active vectors, I0 is the zero vector, ω Projecting the resulting complex vector onto the
is the angular velocity, θ is the angular displacement real and imaginary axes leads to:
and the Cartesian coordinate system represents the
imaginary and real axes. The active vectors form a 𝑰!"# cos θ 𝑇! = 𝐼!" (𝑇! + 𝑇! )
symmetric hexagon with six identical sectors, while ! (5)
the zero vector I0 lies at the center of the hexagon. 𝑰!"# sin θ 𝑇! = 𝐼 −𝑇! + 𝑇! .
! !"

vectors (Fig.3), based on the switching pattern and Solving Eqs.(3) and (5) yields the dwell times T0,
frequency of the inverter. As Iref passes through the T1, and T2 at any given θ angle as shown in Eq.(6):
vectorial spaces, different switch sets are applied to
!
formulate the desired output. As a result, when Iref 𝑇! = 𝑚! sin ( − θ)𝑇!
!
rotates one revolution, the output continues for one !
𝑇! = 𝑚! sin ( + θ)𝑇!
cycle based on the fundamental frequency. ! (6)
𝑇! = 𝑇! − 𝑇! −𝑇!
! !
3.2. Calculating Dwell-Time for − ≤θ≤ ,
! !

The dwell time for stationary vectors represents the time where ma = | Iref | / IDC is the modulation index in the
that space vectors spend in a sector during TS. The basic domain: ma ∈ [0, 1].
method used for dwell-time calculations is the so-called When Iref is inside of other sectors a multiple of
ampere-second balancing principle. The main idea π/3 is subtracted from θ so that the modified angle σ
behind it is that the product of the sampling period TS ! !
happens to fall within the range of − ≤ θ ≤ for use
and the reference vector Iref is equal to the sum of the ! !
current vectors multiplied by the time interval of chosen in Eq.(7):
space vectors. With small enough TS, Iref can be
!
assumed as constant over a period. With this notion, the σ = θ − (𝑘 − 1)
!
reference current is approximated from the 𝑠 = 1, 2, … , 6 (7)
neighbouring vectors corresponding to the actual sector ! !
for − ≤θ≤ ,
and from the zero vector for adjusting the length. ! !
The ampere-second balancing principle is given by
where s is the sector index. This method save sthe need
𝑰!"# 𝑇! = 𝑰! 𝑇! + 𝑰! 𝑇! + 𝑰! 𝑇! for implementing unnecessary calculations and code
(3) capacities.
𝑇! = 𝑇! + 𝑇! + 𝑇!

where T0, T1, and T2 denote the dwell times of the 3.3. Switching Sequence
vectors I0, I1, and I2. From Eq.(2), the individual values
of space vectors in SECTOR I are defined in Eq.(4): As far as designing the switching pattern of the SVM
for the CSI a few notions should be specified for
! ! switching loss reduction:
𝑰! 𝑇! = 𝐼!" e!! ! • The state transition should involve the least
!
!
! (4) amount of switches (one switch is desired).
𝑰! 𝑇! = 𝐼!" e! !
! • Minimum amount of switches should be
𝑰! 𝑇! = 0 involved as Iref transits from the current to the
next state.

44(2) pp. 105–111 (2016) DOI: 10.1515/hjic-2016-0013

Unauthenticated
Download Date | 7/5/17 4:28 AM
108 NEUKIRCHNER AND MAGYAR

Table 2. Typical switching sequence for Sector I. The are needed. Hence, linear differential equations for the
sequence was repeated twice to achieve a smoother !
output signal. In this case, the sampling period TS is
state variable x = 𝑖!" 𝑉!!" 𝑉 !" 𝑖 𝑖 , input variable u =
! !! ! !
one 12th of the fundamental time period. The active VDC and output variable y = [iR iS]T can be derived using
switch pairs are S6, S1 and S1, S2 for space vectors I1 KVL and KCL. Arranging the expressions for the
and I2. The chosen zero state (with switches S1, S4) is aforementioned state variables and recognizing VDC as
chosen for I0 such that the minimum switching loss the input of the system, yields the following state-space
requirement is met. matrices for the zero switching state as shown in
Vectors I1 I2 I0 I1 I2 I0 Eqs.(10)-(12):
S1 X X X X X X
!!"
S2 X X − 0 0 0 0
S3 !!"
! ! !
S4 X 0 − 0 −
!! !! !!
S5
! ! !!
S6 X X X 𝐀𝟎 = 0 0 − − − (10)
!! !! !!
dwell T1 T2 T0 T1 T2 T0 !! ! !!
time TS TS 0 0
!! !! !!
! ! !!
0 0
The typical sequence can be observed in Table 2. !! !! !!
Iref is created by I1, I2, and I0 with the mentioned dwell
! !
time rule 6. Ts can be divided into three or more
𝐁 = −! 0 0 0 0 (11)
segments (T0, T1, and T2) according to the desired THD !"

value of the output current.

0 0 0 1 0
From the above, a continuous switching sequence 𝐂= , (12)
for the SVPWM modulation can be obtained by Eq.(8): 0 0 0 0 1

where Qa = C(3R + Rf), Qb = RRf + 3RRl + RfRl and Qc

𝑰! , 𝑰!!! , 𝑰! 𝑘 = 1, 2, … , 5
(8) =(3R + Rf)(Lf + Ll).
𝑰! , 𝑰! , 𝑰! 𝑘 =6,

where k is the sector index. 3.5. Open-Loop Performance

As CSI exhibits different LTI regions for different state

3.4. Obtaining Open-Loop ‘Piecewise Affine’
matrices A0, A1, ..., A6 according to the used space
Model vector I0, I1, ..., I6 in the corresponding dwell time, the
piecewise affine approach (PWA) can be used to
With every state vector I0, I1, ..., I6 a different switch describe the overall dynamics. The discrete time
configuration and a different circuit configuration is piecewise affine model can be obtained by discrediting
present at a given time. For every case the circuit’s state the set of continuous affine models in the form of
equations can be written in continuous state-space form: Eq.(13)
𝒙(𝑡) = 𝐀 ! 𝒙 𝑡 + 𝐁! 𝒖(𝑡) 𝒖!
𝒚(𝑡) = 𝐂! 𝒙(𝑡) (9) Φ! 𝒙! + 𝚪 𝒖! , if 𝒙! ∈ 𝑃!
s = 0, 1, … , 6 , 𝒙!!! = ⋮ (13)
𝒖!
Φ! 𝒙! + 𝚪 𝒖! if 𝒙! ∈ 𝑃!
where AS is the sth state matrix, BS is the sth input matrix,
CS is the sth output matrix, x(t) and u(t) are space
vectors, and s represents the space vector index with s = Such systems are composed of s local affine
0 as the zero state and s = 1, 2, …, 6 as the active states. models whose parameters in (Φs) change according to
Since all the equation systems have the same input (BS) which polyhedron Pk contains the state vector.
and output (CS) matrices they will be referred to as B The simulated waveforms for VDC = 400 V DC
and C matrices. link with pre-charge resistance RDC = 1 Ω and choke
Any linear circuit can be represented by a number inductance LDC = 2e−1 H using the space vector
of differential equations equal to the number of energy modulation as shown in Fig.5, where iA denotes the
storage units in the circuit. However, in three-phase PWM current of the inverter and iR stands for the phase
balanced circuits, the number of equations can be current through the load. The inverter operates at a
reduced by a third factor because any three-phase fundamental frequency of f = 50 Hz. The filter capacitor
quantity can be expressed in terms of the other two C f = 20e−6 F, the filter inductance Lf = 10e−3 H with
phase quantities. Therefore, since currents flowing filter resistance Rf = 1 Ω and R = 1 Ω representing
through Lf and Ll and voltages across Cf are three-phase losses. The inverter consists of three-phase balanced
balanced quantities only four linear differential inductive load Ll = 2e−3 and resistive load Rl = 5 Ω. The
equations representing the AC side and one linear open-loop performance can be observed in Fig.4 and the
differential equation representing DC current across LDC modulation and PWM formulation in Fig.5.
Hungarian Journal of Industry and Chemistry

Unauthenticated
Download Date | 7/5/17 4:28 AM
MODELLING A THREE-PHASE CURRENT SOURCE INVERTER 109

Figure 4. Three-phase sinusoidal waveform of the CSI output currents iA , iB, and iC with a sample period of Ts = T/36 and
a modulation index of ma = 0.9. The output wave can be smoothened by adding higher filter parameters or further
reducing the sampling period.

Figure 5. Comparison of the inverter output current iA and the filtered load current iR with Ts = T /36 and ma = 0.9. The
background s’s value indicates the corresponding affine state of the model at a given time.

4. MPC assisted SVPWM been applied in the field of power electronics, where the
control frequency ranges from a few hundred Hz up to a
few MHz [5,6]. The main challenge of implementing
The aim of the previous sections was to develop a MPC consists of solving the associated optimization
model for the inverter that describes the switching problem in a bounded time.
behaviour of the dynamics. The obtained piecewise The predicted values are assumed to be constant
affine model gives a good basis for the model predictive after time k. Then, the control problem can be
control technique that is advantageous from two points formulated as follows:
of view: the possibility of introducing constraints to the
state and input variables and the fact that optimal
control problems can be formulated within it. min J (x[k])
The aim is to find the optimal space vectors in s.t. x[k] = x[k|k]
such a way that the error between the predicted states of J(x[k]) = ! !!! ||𝑄(𝒙 𝑘 + 𝑗 𝑘] − xref ([k]))||
2
(14)
the inverter and the chosen reference values are
minimized with respect to constraints. Previously, the where x[k + j | k] denotes the prediction of the state at
method of model predictive control was reserved only time k + j based on the information available at time k,
for applications with relatively slow time scales, like N is the prediction horizon, Q is the constant weighting.
plants and chemical processes. The reason for this is The optimal state-finding algorithm follows:
that MPC requires an optimization problem to be solved
at each sampling instant in order to obtain the control OptimalError = inf
input. Over the past decade, more and more applications for s = 1 : 7 do
use fast MPC applications, thanks to the advances in 𝑥(1) 𝑥(2) … 𝑥(𝑁) ! =
computer technology. The optimization problem is often 𝐵! 0 ⋯ 0 𝑢 0 𝐴!
NP hard, and finding a solution is taxing in terms of the 𝐴! 𝐵! ⋯ 0 𝑢 1 𝐴!!
calculation time of the control device, thus limiting the ∙ + 𝑥(0)
⋮ ⋮ ⋱ ⋮ ⋮ ⋮
rate at which the control could intervene with the !!! !!! !
𝐴! 𝐵! 𝐴! 𝐵! ⋯ 𝐵! 𝑢 𝑁 𝐴!
application. In particular, model predictive control has

44(2) pp. 105–111 (2016) DOI: 10.1515/hjic-2016-0013

Unauthenticated
Download Date | 7/5/17 4:28 AM
110 NEUKIRCHNER AND MAGYAR

Figure 6. MPC reference tracking with f = 50Hz, 𝐼̂ = 37.5A, and N = 2. In order to create the simplest implicit MPC for a
finite state PWA environment, every affine state was tested with the according horizon in every iteration.

Figure 7. Comparison of classic SVPWM and MPC on the same system. The SVPWM is applied with a modulation
index of 0.9, and the MPC reference tracking with f = 50Hz, 𝐼̂ = 37.5A, and N = 2.

Errors = Js(x) modulation. The obtained overall model is in a

if Errors < OptimalError then piecewise affine form that supports the use of model
OptimalError = Errors predictive control in the future. The model has been
Optimal_s = s verified against engineering expectations and its open-
loop performance shows that it is a promising basis of
where OptimalError represents the minimal error value, further work in terms of identification and control. As
Error_s is the actual error value in each iteration, and the next step, a model predictive control synthesis was
Optimal_s is the optimal state. The algorithm uses a performed based on the technology-driven optimality
parallel implicit MPC scheme for every state. Since criteria to assist the open-loop SVPWM modulated
there are only seven states, with a low horizon the system. Due to the implicit MPC-assisted modulation
algorithm can be feasibly implemented. When f = 50Hz, the optimal space vector could be found and utilized in
𝐼 = 37.5A, and N = 2, deviations from the optimal each state over an acceptable timescale due to the low
trajectory are present during optimization as shown in horizon. As such a significantly lower time response
Fig.6, but they are still acceptable in terms of harmonic and better signal shape was achieved.
distortion. The performance observed in Fig.7 can be
compared to the open-loop solution. It can be stated that
due to the rigid structure of the open-loop modulation’s
SYMBOLS
rigid structure the dynamical properties of the MPC-
assisted system exhibit better performance in terms of A state matrix
time response and signal shape. B input matrix
C output matrix
Cf filter capacitance
5. Conclusion
f fundamental frequency
Φ discrete state matrix
A current source inverter model has been developed in Γ discrete input matrix
this paper that is constructed from six LTI models for I0,I1,..., I6 space vector
the different switching modes using SVPWM iA, iB, iC inverter output currents per phase

Hungarian Journal of Industry and Chemistry

Unauthenticated
Download Date | 7/5/17 4:28 AM
MODELLING A THREE-PHASE CURRENT SOURCE INVERTER 111

iDC DC link current Acknowledgement

iR, iS, iT load currents per phase
Iref reference current value
k sector index This research is supported by the National Research
Lf filter inductance Development and Innovation Office (NKFIH) through
Ll load inductance grant No.115694. A. Magyar was supported by the
LTI Linear Time Invariant János Bolyai Research Scholarship of the Hungarian
ma modulation index Academy of Sciences.
MPC Model Predictive Control
νcr triangular carrier wave for TPWM REFERENCES
νm modular wave for TPWM
p discrete time index
P space vector containing polyhedron [1] Espelage, P.; Nowak, J.M.; Walker, L.H.:
PWA ‘Piecewise Affine’ Current Source Inverter Symmetrical GTO current source inverter for wide
(CSI) speed range control of 2300 to 4160 V, 350 to
PWM Pulse Width Modulation 7000 hp induction motors, IEEE Ind. Appl. Soc.
R loss resistance Conf. (IAS) 1988 1, 302–307 DOI 10.1109/IAS.1988.25079
RDC DC link start-up pre-charging resistance [2] Wu, B.: High-power converters and AC drives
Rf filter resistance (Wiley-IEEE Press, New Jersey, USA) 2006
Rl load resistance [3] Ma, J.; Wu, B.; Rizzo, S.: A space vector
s space vector index modulated CSI-based AC drive for multi-motor
S1, S2,..., S6 switching devices applications, IEEE Trans. Power Electron., 1999
σ modified angle 16(4), 535–544 DOI: 10.1109/APEC.1999.750460
SHE Selective Harmonic Elimination [4] Wiseman, J.; Wu, B.; Castle, G.S.P.: A PWM
SVPWM Space Vector Pulse Width Modulation current source rectifier with active damping for
T fundamental cycle time high power medium voltage applications, IEEE
Tk dwell time of corresponding space vectors Power Electron. Spec. Conf. (PESC) 2002 1,
Ts sampling period 1930–1934 DOI: 10.1109/PSEC.2002.1023096
TPWM Trapezoidal Pulse Width Modulation [5] Mariéthoz, S.; Morari, M.: Explicit Model-
θ angular displacement Predictive control of a PWM inverter with an LCL
u input vector filter, IEEE Trans. Ind. Electron. 2008 56(2), 389–
VDC DC link voltage source value 399 DOI: 10.1109/TIE.2008.2008793
𝑉!!"
!
, 𝑉 !"
!! , 𝑉 !"
!! filter capacitance’s voltage
[6] Mariéthoz, S.; Almér, S.; Baja, M.; Beccuti, A.G.;
x state vector Patino, D.; Wernrud, A.; Buisson, J.; Cormerais,
y output matrix H.; Geyer, T.; Fujioka, H.; Jonsson, U.T.; Kao,
ω angular velocity C.Y; Morari, M.; Papafotiou, G.; Rantzer, A.;
Riedinger, P.: Comparison of hybrid control
techniques for buck and boost DC-DC converters,
IEEE Trans. Control Syst. Technol. 2010 18(5),
1126–1145 DOI: 10.1109/TCST.2009.2035306

44(2) pp. 105–111 (2016) DOI: 10.1515/hjic-2016-0013

Unauthenticated
Download Date | 7/5/17 4:28 AM