516 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 21, NO.

2, JUNE 2006

Robust Stability Analysis of Voltage and Current

Control for Distributed Generation Systems
Mohammad N. Marwali, Member, IEEE, Min Dai, Student Member, IEEE, and Ali Keyhani, Fellow, IEEE

Abstract—This paper analyzes the robust stability of a voltage technique does not perform well under nonlinear load, which
and current control solution for a stand-alone distributed genera- significantly undermines its value for power supply applications.
tion (DG) unit using the structured singular value or µ-framework- Lin et al. [5] designed a dc-dc power converter controller using
based method. The voltage and current control solution consists of
a discrete-time sliding mode current controller and a perfect ro- the structured singular value (µ) concept, which evaluates how
bust servomechanism problem voltage controller, which has been stable the system is under the worst case of perturbation. This
shown effective in transient response and harmonic minimization study uses admittance instead of resistance to model the dc load,
in previous study without considering parametric uncertainty. The which is proved convenient in the analysis. However, this design
stability robustness of the system and its transient performance considers only load disturbances and no parametric uncertain-
are investigated under various tuning parameters of the controller.
The analysis results presented in this paper demonstrate that the ties are included in the perturbation. Mohamed [6] proposed a
controller parameters can be tuned and verified to satisfy a certain robust controller for a current source inverter (CSI) fed induction
transient performance requirement and at the same time guaran- motor drive. Both H∞ loop-shaping and µ-analysis techniques
tee robust stability under system parameter uncertainties and load are applied in the research, but no parametric uncertainty is con-
variations. sidered, which undermines the strength. Ye et al. [7] proposed a
Index Terms—Digital signal processors (DSP), pulse width mod- robust controller design method for high frequency resonant in-
ulated (PWM) inverters, robust stability, structured singular value. verters. This approach applies the H∞ robust controller synthe-
sis method provided in the Matlab Robust Control Toolbox but
I. INTRODUCTION includes only load and external input voltage in the perturbation.
FEEDBACK control system is said to achieve robust sta- A perfect robust servomechanism problem (RSP) controller
A bility if it remains stable for all considered perturbations in
the plant. The stability robustness of the system is evaluated by
was developed in [8], which guarantees exact asymptotic track-
ing of the fundamental frequency reference and error regulation
its tolerance to perturbations. In feedback-controlled pulsewidth of the load disturbance at each harmonic frequency included in
modulation (PWM) inverter systems, e.g., an inverter-based the servo compensators. The perfect RSP guarantees this prop-
three-phase distributed generation (DG) unit operated in stand- erty independent of any perturbations in the plant as long as
alone mode, load disturbance, noise, and parametric uncertainty they do not destabilize the system. The perfect RSP guarantees
of the electrical components in the circuit are the major plant per- stability under a nominal plant without perturbation; however,
turbations that have significant impacts on both system stability the stability under perturbation is not guaranteed and this issue
and performance and therefore warrant detailed investigation. is not addressed in [8]. Therefore, it is important to analyze the
Robust stability-related topics about PWM inverter-based sys- stability property of the controller under possible disturbances
tems have been addressed in literature. Czarkowski et al. [1] in order to ensure proper operation of the converter over its
studied a state feedback control method of a PWM dc-dc con- intended operating range.
verter for its robust stability under parametric uncertainty. This In this paper, the stability robustness of the system with the
study used Kharitonov’s theorem [2], which checks whether the controller developed in [8] will be investigated using structured
feedback system is stable by applying the Routh–Hurwitz sta- singular values or a µ-framework. Specifically, perturbations
bility tests but does not tell the stability margin or how stable the due to load variations and parameters uncertainties of the sys-
system is. Grundling et al. [3] developed a robust model refer- tem components are considered. A linear quadratic cost function
ence adaptive control technique for uninterruptible power sup- with separate weighting scalars for plant states and servocom-
plies (UPS) which was expected to handle model inaccuracy, but pensator states have been used to find solutions to the perfect
no robust stability property of the technique was presented. Lee RSP. In this paper, the stability robustness and transient response
et al. [4] proposed an H∞ loop-shaping robust controller design of the resulting control system will be investigated for different
technique for UPS with robust stability analysis. However, this choices of these weighting scalars. The transient performance
of the system is evaluated by performing moving window RMS
Manuscript received August 7, 2005; revised August 7, 2005. This work calculations of the three-phase output voltages under transient
was supported in part by the National Science Foundation under Grants ECS load change from zero to 100% resistive load.
0501349 and ECS 0105320. Paper no. TEC-00147-2005.
M. N. Marwali is with the Liebert Corporation, Delaware, OH 43015 USA
A review of the voltages and current controller developed
(e-mail: nanda marwali@liebert.com). in [8] is first presented and followed by a summary of robust sta-
M. Dai and A. Keyhani are with the Department of Electrical and Com- bility theory using a structured singular value or µ-framework.
puter Engineering, Ohio State University, Columbus, OH 43210 USA (e-mail:
dai.21@osu.edu; keyhani.1@osu.edu).
The uncertainty model is then developed and used for verifying
Digital Object Identifier 10.1109/TEC.2005.860406 the robust stability of the system.

0885-8969/$20.00 © 2006 IEEE

MARWALI et al.: ROBUST STABILITY ANALYSIS OF VOLTAGE AND CURRENT CONTROL FOR DISTRIBUTED GENERATION SYSTEMS 517

Fig. 1. PWM inverter based DG system in stand-alone mode.

II. VOLTAGES AND CURRENTS CONTROL The per-unit values of the capacitances, inductances, and resis-
tances are calculated from
Fig. 1 shows a PWM inverter used for the DG system, where
a constant dc voltage source is used to approximate most typical
xlinv ≡ ωf Linv /Z1, xrinv ≡ Rinv /Z1,
DG sources, such as fuel cell, photovoltaic, wind, and micro-
turbine generation systems. This approximation is surely rea- xcinv ≡ 1/(ωf 3Cinv )/Z1,
sonable when these sources work together with a stiff voltage
xlT ≡ ωf LT /Z2, xrT ≡ RT /Z2,
regulation and secondary energy storage when necessary, such
as the systems shown in [9] and [10]. The approximation is also xcload ≡ 1/(ωf Cload )/Z2
reasonable even with an unregulated DC bus due to the adapt-
ability of the PWM inverter as long as the DC voltage is high where the base impedances Z1 and Z2 are defined as
enough to generate the control command.
The DG system shown in Fig. 1, if operated in stand-alone Z1 ≡ V 1/I1 Z2 ≡ V 2/I2
mode, is typically used as stand-alone on-site power or standby
where V 1 and V 2 denote the rated line-to-neutral primary and
emergency power when the utility grid is not available or the
secondary transformer voltages, and I1 and I2 denote the rated
utility power is accidentally lost due to fault.
primary and secondary currents, respectively. The voltages and
A state space model of the system in DQ0 stationary reference
currents in (1) are the DQ0 stationary reference frame variables
using per-unit notation can be developed and is given by (1) [11].
of the per-unit voltages and currents in ABC given as

pwmabc /V 1,
invabc /V 1
d
v invq d /dt = (
iinvq d − ·
isndq d )/cinv (1.a)
v pwmabc = V
v invabc = V

abc /I1,

iinvabc = Iinv
loadabc /V 2

v loadabc = V
d
iinvq d /dt = (−rinv · īinvq d −
v invq d +
v pwmq d )/linv



iloadabc = Iloadabc /I1
(1.b)
d
v loadq d /dt = (
isndq d −
iloadq d )/cload (1.c) where the actual ABC voltages and currents are defined as shown
in the equation at the bottom of the next page. The following
d
isndq d /dt = (−rT ·
isndq d +
v invq d −
v loadq d )/lT changes of variables have been used for the secondary trans-
(1.d) former currents, load currents, and load voltages in (1):
dvload0 /dt = (isnd0 − iload0 )/cload (1.e)
isndq d = triq d ·
isndq d (2.a)
disnd0 /dt = (−rT · isnd0 − vload0 )/cload (1.f)
iloadq d = triq d ·
iloadq d (2.b)

v loadq d = trvq−1
d ·

v loadq d (2.c)
where the following per-unit capacitances, inductances, and re-
sistances are defined from their corresponding per-unit values: with matrices triq d and trvq d defined as
 

 −1 0
cinv ≡ 1/(ωf · xcinv ); cload ≡ 1/(ωf · xcload ); triq d = Ks · tri · KS−1 row1, 2 = (3.a)
col1, 2 0 −1
linv ≡ xlinv /ωf ; ltrans ≡ xlT /ωf ; rinv ≡ xrinv ;  
−1 0
trvq d = [Ks · trv · KS−1 ] row1, 2 = . (3.b)
rT ≡ xrT . col1, 2 0 −1
518 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 21, NO. 2, JUNE 2006

The RSP is adopted for voltage control due to its capabil-

ity to perform zero steady-state tracking error under unknown
load and eliminate harmonics of any specified frequencies with
guaranteed system stability for nominal plant parameters and
load variations. The theory behind the RSP is based on the so-
lution of the RSP [12] where the internal model principle [13]
and the optimal control theory for linear systems are combined.
The RSP controller consists of a discrete form of the continuous
Fig. 2. The RSP and DSM controllers. servocompensator:

η˙ = Ac
η + Bc eV , eV = V ref − V load (6)
Matrices tri and trv in (3) denote currents and voltages trans-
formations of a particular delta-wye transformer, e.g., where
   
η = [
η1 ,
η2 , . . .
ηn ]T
ηi ∈ R2 , i = 1, 2, . . . n
−2 1 1 −1 0 0
1
tri = 1 −2 1  , trv =  0 −1 0  . Ac = block diag[Ac1 , Ac2 , . . . , Acn ]
3
1 1 −2 0 0 −1 Bc = [Bc1 , Bc2 , . . . , Bcn ]T
From (1) it can be seen that the zero variables are not affected with
by the control inputs and therefore need not be considered in the
0 1
controller design. Moreover, the D and Q axes are completely Aci = , i = 1, 2, . . . n
−ωi2 0
decoupled and have the same dynamics. Therefore, a controller
can be developed and analyzed for one of the dq axes and the Bci = (0 1)T , i = 1, 2, . . . n.
same controller can be used for the other.
Fig. 2 shows the voltages and currents control developed Note that each of the blocks
η˙ i = Ac1
ηi + Bci
eV q d repre-
in [8]. An RSP controller is used for the voltage control, and sents a state space implementation of the continuous trans-
a discrete-time sliding mode (DSM) controller is used for the fer function 1/(s2 + ωi2 ), where each ωi = 2πfi represents
current control. the fundamental frequency to track and the harmonic frequen-
The DSM controller is used in the current loop to limit the cies to be eliminated. For a 60-Hz DG system with desire to
inverter current under overload conditions because of the fast eliminate fifth and seventh harmonics, for example, we use
and no-overshoot response it provides. Suppose a discrete form ω1 = 2π · 60, ω2 = 2π · 5 · 60, and ω3 = 2π · 7 · 60.
of the LC filter dynamics given in (1.a) and (1b) is given by To design the voltage controller using the perfect RSP, we
need to consider a combination of the true plant in (1) and the

x1 (k + 1) = A∗1
x1 (k) + B1∗
u(k) + E1∗ d
1 (k) (4) DSM current controller (5) as the equivalent “plant” seen by
the outer voltage loop. Assuming a discrete form of (1) given
where the states are
x1 = [
v inv,
iinv], the inputs u = vpwm, by
x∗p (k + 1) = A∗p
x∗p (k) + Bp∗ vpwm(k), the augmented true
and disturbances d
1 = isnd. Note the subscript dq has been plant and DSM current controller can be found as in
dropped since the controller is designed for one axis but applied

x∗p (k + 1) = Ad
x∗p (k) + Bd · icmd(k) (7)
for both. To force the inverter currents to follow their commands,
the sliding mode surface is chosen as s(k) = C1 ·
x1 (k + 1) − with
icmd(k), where C1 ·
x1 (k + 1) = iinv(k + 1), so that when
Ad = A∗p − Bp∗ (C1 B1∗ )−1 (B1∗ C11 + E1∗ C12 )
DSM occurs, we have
s(k) = 0 or iinv(k + 1) = icmd(k). The
existence of the DSM can then be guaranteed if the control is Bd = Bp∗ (C1 B1∗ )−1
given by  
1 0 0 0 0
C11 =
vpwm(k) = (C1 B1∗ )−1 (
icmd − C1 A∗1
x1 (k) − C1 E1∗ d
1 (k)). 0 1 0 0 0
(5) C12 = [0 0 0 1 0].

√

pwmabc = [V pwmab
V V pwmbc V pwmca ]T / 3
√

invabc = [V invab
V V invbc V invca ]T / 3

loadabc = [V loada
V V loadb V loadc ]T ,


Iloadabc = [Iloada Iloadb Iloadc ]T


Isndabc = [Isnda Isndb Isndc ]T ,
√

abc = [Iinva − Iinvb ,
Iinv Iinvb − Iinvc , Iinvc − Iinva ]T / 3
MARWALI et al.: ROBUST STABILITY ANALYSIS OF VOLTAGE AND CURRENT CONTROL FOR DISTRIBUTED GENERATION SYSTEMS 519

The complete perfect RSP voltage controller is given by states

feedback
icmd(k) = K0 x∗p (k) + K1 η(k) (8)
where the gains K = [K0 K1 ] are found by minimizing a certain
linear quadratic cost function for the augmented “equivalent
plant” (7) and a discrete form of the servocompensator (6):
 ∗    ∗ 

xp (k + 1) Ad 0
xp (k)
=
η(k + 1) −Bc∗ C A∗c η(k)
 
Bd
+ u (k). (9)
−Bc∗ D 1

To achieve the desired transient performance and guarantee

robust stability under plant uncertainties, the following linear
quadratic cost function has been used in this paper:
∞
wp · x∗p (k) · x∗p (k) + wS 1 · η1 (k) · η1 (k)
Jε =
+wSH h ηh (k) · ηh (k) + ε · u(k) u(k)
k =0

(10)
where wp , wS , and wSH represent weighting scalars for plant
states (x∗p ), fundamental servocompensator states (η1 ), and har-
monics servocompensator states (ηh ), respectively. Solution to
this linear quadratic optimization problem is well known and
can be found using Matlab command dlqr [14].
The effectiveness of the technique can be demonstrated by ex-
perimental results shown in Fig. 3, which exhibits well-regulated
sinusoidal output voltage waveforms under various types of
load.

III. ROBUST STABILITY ANALYSIS USING STRUCTURED

SINGULAR VALUE µ
Structured singular value µ can be used to analyze and
evaluate the stability robustness of a multi-input multi-output
(MIMO) linear system under structured perturbations. In order
to use the µ-framework to analyze the robust stability of a lin-
ear system under perturbation, the problem needs to be recast
into a feedback loop diagram of Fig. 3, where M represents a
known stable MIMO transfer function of the linear system with
n inputs and n outputs and a structured uncertainty matrix of
the form (11) [15], [16]:
∆ = {diag[δ1 Ir 1 , . . . δs Irs , ∆1 . . . ∆F ] : δi ∈ C, Fig. 3. Experimental results of the stand-alone DG system under differ-
ent scenarios, where the top traces are three-phase output voltages and the
∆j ∈ C m j xm j } (11) bottom traces are three-phase load currents: (a) transient response at load
stepping up from 0 to 100%; (b) transient response at load stepping down
where Si=1 ri + Fj=1 mj = n. from 100% to 0; (c) steady-state performance under three-phase nonlinear
load.
The structured singular value of M with respect to the uncer-
tainty set ∆ is defined as
1
µ∆ (M ) = . A. Uncertain Open-Loop Model
min{σ̄(∆) : ∆ ∈ ∆, det(I − M ∆) = 0}
In order to use the µ-framework to analyze robust stability,
The generalized small-gain theorem provides robust stability the problem needs to be recast to that of Fig. 4. A class of general
results of the system using the structured singular value. It states feedback loops called linear fractional transformations (LFTs)
that if nominal M (s) is stable then the perturbed system (I − can be used to achieve this. For complete discussion of LFTs
M ∆)−1 is stable for all stable ∆i for which ∆i ∞ ≤ 1 if and and their use in representing model uncertainties, refer to [15].
only if µ∆ (M (jω)) < 1 for all ω ∈ R [15], [16]. Using the state space model (1), a single-phase equivalent circuit
520 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 21, NO. 2, JUNE 2006

Fig. 4. Representation of a linear system with uncertainties.

Fig. 6. LFTs for (a) 1/cinv -inverse term and (b) rinv -noninverse term.
Fig. 5. Open-loop model of the nominal plant.

gload = gload0 (1 + τgload · δgload ), |δgload | < 1 (13.g)

of the converter with RL load can be derived as shown in Fig. 5.
This represents the open-loop model of the plant. λload = λload0 (1 + τλload · δλload ), |δλload | < 1 (13.h)
The dynamic equations of the plant are given by
where the terms with subscript zero indicate the nominal values
dvinv/dt = (iinv − ·isnd )/cinv (12.a) of the parameters and τ terms denote their percentage tolerances
given above. The following load parameters can be used to rep-
diinv/dt = (−rinv · iinv − vinv + vpwm)/linv (12.b)
resent the desired load variation above: gload0 = 0.8, λload0 =
dvload/dt = (isndsnd − gload vload − iL )/cload (12.c) 0.6ωf with τgload = τλload = 1.0.
   Parameters cinv , linv , cload , and lT appear in denominators in
disnd /dt = (−rT · isnd + vinv − vload )/lT (12.d)
(12) and can be represented using a lower LFT similar to that
diL /dt = λload vload (12.e) shown in Fig. 6(a) for 1/cinv . LFTs for rinv , rT , rload , and λload
are constructed similar to that of Fig. 6(b) for rinv , where
iload = 
gload vload + iL (12.f)
drinv = rinv0 · τrinv (14.a)
where gload = 1/rload is the per-unit conductance of the load
and λload = 1/lload is the inverse of the per-unit load induc- drT = rT 0 · τr T (14.b)
tance.
dgload = gload0 · τgload and (14.c)
Let us assume that the following parameter variations exist in
the system due to manufacturing tolerances of the components dλload = λload0 · τλload . (14.d)
used and/or errors in the parameters identification processes:
1) inverter filter capacitor tolerance: ±6%; For each uncertain term, the corresponding LFT as shown
2) inverter filter inductor tolerance: ±15%; in Fig. 6 can be substituted into the dynamic model of (12) as
3) inverter filter inductor losses tolerance: ±50%; shown in Fig. 7, with the uncertain perturbation (δ) blocks sep-
4) output filter capacitor tolerance: ±6%; arately represented by ∆(s) as shown in Fig. 8, where ∆(s) is
5) transformer filter inductor tolerance: ±15%; given by {δCinv , δLinv , δCload , δLT , δRinv , δRT , δGload , δΓload }.
6) transformer filter inductor losses tolerance: ±50%. The block P (s) in Fig. 8 contains all the known model informa-
Furthermore, the specification of the unit requires that it op- tion of the plant and will be referred to as the nominal open-loop
erates stably without degradation in performance for load from plant model. The combination of P (s) and ∆(s) forms the un-
zero to 200% with a power factor of 0.8 lagging at maximum certain open-loop model.
load. This specification represents variation of the resistive load The state space model of the nominal plant P can be derived
from zero to 160%, and inductive load from zero to 120%. by inspection of Fig. 7 with the following states, inputs, and
The above parameter variations due to manufacturing toler- outputs variables as shown the equations at the bottom of the
ances and load variations can be precisely written as follows: page 7. The state space equations are given as

cinv = cinv0 (1 + τcinv · δcinv ), |δcinv | < 1 (13.a) Ẋ = AP X + BP U

linv = linv0 (1 + τlinv · δlinv ), |δlinv | < 1 (13.b) Y = CP X + DP U (15)

cload = cload0 (1 + τcload · δcload ), |δcload | < 1 (13.c)

where
lT = lT 0 (1 + τlT · δlT ), |δlT | < 1 (13.d)
AP = [Anom ], BP = [Bnom Bdel ],
rinv = rinv0 (1 + τrinv · δrinv ), |δrinv | < 1 (13.e)    
Cnom Dnom
CP = , DP = .
rT = rT 0 (1 + τr T · δr T ), |δr T | < 1 (13.f) Cdel Ddel
MARWALI et al.: ROBUST STABILITY ANALYSIS OF VOLTAGE AND CURRENT CONTROL FOR DISTRIBUTED GENERATION SYSTEMS 521

Fig. 7. Block diagram of the nominal plant P (s) with LFTs of the uncertain parameters.

The terms with subscription nom are the nominal model param- Control System Toolbox [14] and µ-Analysis and Synthesis
eters derived from (12), and the terms with subscription del are Toolbox [15] have been utilized to achieve this purpose.
defined in (16), (17), and (18), respectively, shown at the bottom Since the controller is implemented in the discrete time sys-
of the next page. tem, the following steps are necessary to obtain the closed-loop
plant model:
B. Uncertain Closed-Loop Model 1) Apply a zero-order hold transformation to the continu-
In order to use the µ-framework to analyze the robust stability ous plant P , to include the effect of the sample and hold
of the system, the system needs to be recast into that of Fig. 4. In process of the digital sampling process.
this case, the system M comprises the nominal open-loop plant 2) Transform the discretized plant back to a continu-
P (s) and the controller loop closed around it as illustrated in oussystem in the w-plane by applying an inverse Tustin
Fig. 8, with the vectors X, W , and Z as defined above. Matlab’s transformation. This transformation has the property of
522 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 21, NO. 2, JUNE 2006

4) The combined controller system is then transformed into

the w-plane in continuous domain using the inverse Tustin
transformation.
5) Finally, the closed-loop plant model M is obtained by
invoking the Matlab sysic command for the transformed
nominal plant model from step 2 and the controller
system obtained in step 4, as shown in the equation at the
bottom of the page.

IV. TUNING THE CONTROLLER PERFORMANCE

Fig. 8. Uncertain closed-loop model.
Recall earlier that the wp , wS , and wSH in cost function
(10) are the weighting scalars for the plant, fundamental, and
harmonic servocompensator states, respectively. The selection
preserving the frequency response of the discrete time of values of these scalars provides a way of tuning the controller
systems. for desired transient performance and stability robustness. The
3) Obtain state-space representations of the RSP and DSM structured singular value discussed in the previous sections
controller with inputs and outputs definitions as shown in will be used to evaluate the robust stability of each controller
Fig. 2. The combined controller state-space system can resulting from different choices of weighting scalars. A time
be calculated using the sysic (µ-Toolbox) command in response simulation of the single-phase equivalent circuit
Matlab [15]. will be used to compare the transient performance of each

 
−τcinv 0 0 0 0 0 0 0
−dr inv
 0 −τlinv 0 0 0 0 0 
 l inv0 
Bdel 
= 0 0 −τcload 0 0 0 −dg load
0  (16)
c load0 
 0 0 0 −τlT 0 −dr T
0 0 
lT 0
0 0 0 0 0 0 0 dλloadl
 1 −1 
0 c inv0 0 c inv0 0
 −1 −r inv 0
0 0 0 
 l inv0 l inv0 
 0 −g load0 1 −1 
 0 c load0 c load0 c load0 
 1 −1 −r T 0 
Cdel 
=  lT 0 0 0  (17)
rT 0 rT 0 
 0 1 0 0 0
 
 0 0 0 1 0
 
0 0 1 0 0
0 0 1 0 0
 
0 −τcinv 0 0 0 0 0 0 0
−dr inv
 l inv0
1
0 −τlinv 0 0 0 0 0
 l inv0 
 0 0 0 −τ 0 0 −dg load
0 0
 cload c load0 
 −dr T 
Ddel = 0 0 0 0 −τ lT 0 0 0 (18)
 lT 0 
 0 0 0 0 0 0 0 0 0
 
 0 0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0

States: X = [vinv iinv vload isnd iL ]T .

Inputs: U = [vpwm W T ]T , where
W = [wcinv wlinv wcload wlT wrinv wr T wgload wλload ]T .
Outputs: Y = [vinv iinv vload iload Z T ]T , where
Z = [zcinv zlinv zcload zlT zrinv zr T zgload zλload ]T .
MARWALI et al.: ROBUST STABILITY ANALYSIS OF VOLTAGE AND CURRENT CONTROL FOR DISTRIBUTED GENERATION SYSTEMS 523

Fig. 9. Transient response for Case 1: w p = 0.5, w s = 5 × 105 , and w SH =

w S . The top trace is the output voltage and its reference, the middle trace is the
load current, and the bottom trace is RMS variations of the output voltage, all Fig. 11. Upper bound of the structured singular values for different w p , and
versus time in seconds. w SH = w S .

Fig. 12. RMS output voltage variations during 0 to 100% and 100% to 0 for
different w p , and w SH = 0.01 × w S .

Fig. 10. RMS output voltage variations during 0 to 100% and 100% to 0 for
different w p , and w SH = w S .

controller. To quantify the transient performance, it is common

in the industry to use the deviation of the RMS output voltage
from its nominal value during a 100% resistive load change
as a performance measure. A less than 5% deviation of
output voltage under 100% resistive load transient is not an
uncommon specification in the industry for a high performance
DG unit.
Consider the following cases of weighting scalars.
1) Case 1: wp = 0.5, ws = 5 × 105 , and wSH = wS
2) Case 2: wp = 0.1, ws = 5 × 105 , and wSH = wS
3) Case 3: wp = 0.05, ws = 5 × 105 , and wSH = wS
4) Case 4: wp = 0.005, ws = 5 × 105 , and wSH = wS
5) Case 5: wp = 0.0001, ws = 5 × 105 , and wSH = wS .
In all the cases above, equal weighting scalars are applied Fig. 13. Upper bound of the structured singular values for different w p , and
to the fundamental compensator states and the harmonic states w SH = 0.01 × w S .
524 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 21, NO. 2, JUNE 2006

Fig. 14. Individual-perturbation frequency response for system with controller in Case 9 (X-axis in each plot is frequency in rad/s).

(wSH = wS ), while the plant states weighting wp in each case in allowing the plant states to move more freely and hence
are decreased from 0.5 all the way down to 0.0001. The time produces a faster response. Notice, however, that only Cases 4
response simulations for Case 1 are illustrated in Fig. 9 and and 5 result in RMS variations of less than 5%.
show the output voltage, reference voltage, load current, and To analyze the robust stability, the upper bound of the struc-
RMS variation of the output voltage during both 0 to 100% tured singular value in each case is plotted in Fig. 11. It can be
and 100 to 0% resistive load transients. From Fig. 9, it can be seen that only Cases 1, 2, and 3 achieve robust stability under
seen that the output voltage RMS deviates as much as close to the considered structured perturbations with the peak value of
20% for Case 1. µ∆ (M (jω)) in each case being less than 1. Cases 4 and 5—
The time response similar to Fig. 9 was obtained for each of the only cases with acceptable transient performances—do not
Cases 1 to 5, and the resulting RMS output voltage variations are achieve robust stability, with the peak values of µ∆ (M (jω))
plotted collectively in Fig. 10. It can be seen that as the scalar being 1.1 and 4, respectively.
weighting wp decreases, the transient performance improves, Consider now the following cases of weighting scalars.
with the RMS variation as little as 2% for the case of wp = 1) Case 6: wp = 0.5, ws = 5 × 105 , and wSH = 0.01 ×wS
0.0001. These results are not unexpected since the weighting 2) Case 7: wp = 0.1, ws = 5 × 105 , and wSH = 0.01 × wS
scalars represent the penalty applied to each state in the system. 3) Case 8: wp = 0.05, ws = 5 × 105 , and wSH = 0.01 × wS
Intuitively, decreasing wp while keeping the compensator states 4) Case 9: wp = 0.005, ws = 5 × 105 , and wSH = 0.01 × wS
weighting the same decreases the penalty cost applied to the 5) Case 10: wp = 0.0001, ws = 5 × 105 , and wSH =
plant states in the cost function to be minimized. This results 0.01 × wS .
MARWALI et al.: ROBUST STABILITY ANALYSIS OF VOLTAGE AND CURRENT CONTROL FOR DISTRIBUTED GENERATION SYSTEMS 525

V. CONCLUSION
The stability robustness of the control system proposed in [8]
was verified using the structured singular value method under
structured perturbations due to component parameters errors
and linear load variations. It was shown that the scalar weighting
in the optimal control cost function provides a way of tuning
the transient performance of the controller while maintaining
stability robustness of the system under perturbations due to
plant parameter uncertainties.

REFERENCES
[1] D. Czarkowski, L. R. Pujara, and M. K. Kazimierczuk, “Robust stability
of state-feedback control of PWM DC-DC push-pull converter,” IEEE
Trans. Ind. Electron., vol. 42, no. 1, pp. 108–111, Feb. 1995.
Fig. 15. Pole (x)-zero (o) map of closed-loop uncertain system for [2] H. Ozbay, Introduction to Feedback Control Theory. New York: CRC
different values of individual-perturbation in inverter filter inductance Press, 1999.
parameter. [3] H. A. Grundling, E. G. Carati, and J. R. Pinheiro, “Analysis and imple-
mentation of a modified robust model reference adaptive control with
repetitive controller for UPS applications,” in Proc. IEEE IECON, 1998,
In Cases 6–10, wP is assigned the same value as in Cases pp. 391–395.
[4] T. S. Lee, K. S. Tzeng, and M. S. Chong, “Robust controller design for a
1–5, respectively, but smaller weighting is used for the har- single-phase UPS inverter using µ-synthesis,” IEE Proc.—Electr. Power
monic compensator states as compared to the fundamental Applicat., vol. 151, no. 3, pp. 334–340, May 2004.
states. Fig. 12 shows the RMS variations for all cases. Com- [5] J. L. Lin and S. J. Chen, “µ-based controller design for a DC-DC switching
power converter with line and load variations,” in Proc. IEEE IECON,
paring Fig. 10 and Fig. 12, it can be seen that the transient 1996, pp. 1029–1034.
response improves for each of the last five cases, with Cases [6] A. M. Mohamed, “Modern robust control of a CSI-fed induction mo-
9 and 10 now resulting in only 2 and 1% RMS variations, tor drive system,” in Proc. 1998 Amer. Control Conf., 1998, pp. 3803–
3808.
respectively. [7] Z. M. Ye, P. K. Jain, and P. C. Sen, “Robust controller design for high
Fig. 13 shows the upper bound of the structured singular value frequency resonant inverter system with voltage mode control,” in Proc.
in Cases 6–10. It can be seen that reducing wSH improves the IEEE IECON, 2004, pp. 41–46.
[8] M. N. Marwali and A. Keyhani, “Control of distributed generation
stability robustness of each of the cases previously considered. systems—Part I: Voltages and currents control,” IEEE Trans. Power Elec-
The counterpart of Case 4 (Case 9) now achieves robust stability tron., vol. 19, no. 6, pp. 1541–1550, Nov. 2004.
with the µ∆ (M (jω)) peak value of 0.7 giving a robust stability [9] M. Dai, A. Keyhani, J. W. Jung, and A. B. Proca, “A low cost fuel cell
drive system for electrical vehicles,” in Global Powertrain Congress’03
margin of (1/0.7) > 1. Case 10 still does not achieve robust Conf. & Expo., Ann Arbor, MI: Advanced propulsion strategy, Sep. 2003,
stability, with a peak value of µ∆ (M (jω)) still greater than 1. pp. 22–26.
Comparing all the above cases, it can be seen that Case 9 gives [10] R. Teodorescu and F. Blaabjerg, “Flexible control of small wind turbines
with grid failure detection operating in stand-alone and grid-connected
the best transient performance while still maintaining the stabil- mode,” IEEE Trans. Power Electron., vol. 19, no. 5, pp. 1323–1332, Sep.
ity robustness of the system under the structured perturbations 2004.
considered. [11] M. N. Marwali, “Digital control of PWM inverters for high performance
UPS,” Ph.D. dissertation, Dept. Elect. Eng., Ohio State Univ., Columbus,
Continuing with the analysis, it is instructive to see how each OH, Dec. 2004.
parameter perturbation affects the stability robustness of the [12] E. J. Davison and B. Scherzinger, “Perfect control of the robust servomech-
system. For this purpose, the frequency response of each in- anism problem,” IEEE Trans. Autom. Control, vol. 32, no. 8, pp. 689–702,
Aug. 1987.
dividual perturbation can be plotted as shown in Fig. 13 for [13] B. A. Francis and W. M. Wonham, “The internal model principle for linear
the system with the controller in Case 9. It can be seen from multivariable regulators,” Applied Mathematics and Optimization, vol. 2,
Fig. 14 that for all the perturbations, the frequency response no. 2, pp. 170–194, 1975.
[14] Control System Toolbox User’s Guide Version 6 Natick, MA, The Math-
values are less than 1.0, which confirms the stability robustness works, Inc., Jun. 2004.
result of the structured singular value presented earlier. The in- [15] G. J. Balas, J. C. Doyle, K. Glover, A. Packard, and R. Smith, The µ
verter filter inductor has the highest peak frequency response Analysis and Synthesis Toolbox. Natick, MA: Mathworks, 1991.
[16] R. Chiang and M. Safonov, Robust Control Toolbox. Natick, MA: Math-
value, with a peak value of around 0.7, which means that an in- works, 1998.
dividual perturbation in this parameter will be the closest to the
margin of making the system unstable. A closed-loop pole-zero
map of the uncertain closed-loop system in Fig. 8 with all other Mohammad N. Marwali (M’98) received the
B.S.E.E. degree from the Institut Teknologi Bandung,
δ’s set to zero except for the δlinv can be obtained for different Indonesia, in 1993 and the M.S.E.E. and Ph.D. de-
values of |δlinv | less than 1 and greater than 1. This is illustrated grees from the Ohio State University, Columbus, OH,
in Fig. 15. The system is stable at the nominal inductance value in 1997 and 2004, respectively.
He is now with the Liebert Corporation in
(δlinv = 0), δlinv = −0.5, and at the lowest value of the com- Delaware, OH. His research interests are in the area of
ponent tolerance (δlinv = −1.0). As the value δlinv is decreased control of uninterruptible power supplies, distributed
further, the system finally becomes unstable at δlinv , slightly generation systems, electric motor drives, and power
converters.
less than −1.4.
526 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 21, NO. 2, JUNE 2006

Min Dai (S’99) received the B.S. and M.S. degrees Ali Keyhani (S’72–M’76–SM’89–F’98) was with
in electrical engineering from Tsinghua University, the Hewlett-Packard Company and TRW Control
Beijing, China, in 1994 and 1997, respectively, and from 1967 to 1972. Currently, he is a Professor of
the M.S. degree in computer science from the Univer- electrical engineering at the Ohio State University
sity of Alabama, Huntsville, in 1999. He is currently (OSU), Columbus. He serves as the Director of the
working toward the Ph.D. degree at the Department OSU Electromechanical and Mechatronic Systems
of Electrical Engineering, the Ohio State University, Laboratory. His research activities focus on the con-
Columbus. trol of distributed energy systems, design and model-
His research interests include power converter ing of electric machines, control and design of power
control for UPS and distributed generation applica- electronic systems, DSP-based virtual test beds for
tions, fuel cell power conversion, electrical machines, design, control of power systems, automotive sys-
and drives. tems, modeling, parameter estimation, and failure detection systems.
Dr. Keyhani received the OSU College of Engineering Research Award
for 1989, 1999, and 2003. He was an Editor for the IEEE TRANSACTIONS ON
ENERGY CONVERSION and past Chairman of the Electric Machinery Committee,
IEEE Power Engineering Society.