1

A New Self-Healing DSTATCOM for Volt-VAR

and Power Quality Control
Bala Boyi Bukata, Student Member, IEEE, and Yun Li, Fellow, IEEE
AbstractA self-healing scheme integrable into the smart
distribution grid through the custom power devices (CPD) is
proposed. The distribution static compensator (DSTATCOM)
is one such device adorned with an insulated gate bipolar
transistor (IGBT) commutative characteristics for delivering
required reactive compensation at the distribution network.
However, its computer-aided-design (CAD) based mitigation
procedures has recently revealed inadequacies in taming
emerging power quality (PQ) problems at the point of
common coupling (PCC). Further accommodation of virtual
power plants (VPP) envisaged during smart grid transforma-
tion would increase uncertainties and boosts concealment of
the systems gradient information. This paper addresses the
problem by applying model-free predictive control (MFPC)
and describing functions in nonlinear modelling of DSTAT-
COM via a simplex optimisation approach to online tuning.
The new SH-MFPC offers the advantage of bypassing the
computationally over-burdened state-space model and can
sufciently handle constraints, while arresting voltage quality
violations online at high speeds against changes in customers
loads. Experimental SimPowerSystems model has been de-
veloped to compare the effectiveness and efciency of the
novel method against conventional optimisation techniques
within the Matlab/Simulink environment.
Index Termsdistribution network, DSTATCOM, fuzzy
logic control, model-free predictive control, power quality,
simplex algorithm.
I. INTRODUCTION
With the more environmentally friendly smart grid
initiatives during the past few years, intelligent operation
and optimisation of the electricity distribution system have
received an increasing attention in power system research
worldwide. Power ow from the distribution substation to
the customer can be optimised at Volt-Ampere-Reactive
(VAR) level by reducing the reactive power. This is
achievable by re-engineering the look-ahead mechanism
of the model predictive controller (MPC) whilst using
its power to solve online optimal control problem in the
open-loop. Closed-loop performance is achieved through
measurements in the next time step to update the optimi-
sation problem of the linearised model. Earlier successful
Manuscript received October 13, 2012; revised November 30, 2012.
This work was supported in part by the Petroleum Technology Develop-
ment Fund (PTDF). Plot 672, Port Harcourt Crescent, Off Gimbia Street,
Off Ahmadu Bello Way, Area 11, Garki, Abuja, Nigeria.
B. B. Bukata is a PhD student in the School of Engineering, University
of Glasgow, Rankine Building, Oakeld Avenue, Glasgow G12 8LT, U.K.
(email: [email protected])
Yun Li is with the Systems, Power and Energy Research Division,
School of Engineering, University of Glasgow, Rankine Building, Oak-
eld Avenue, Glasgow G12 8LT, U.K. Tel. +44 141 330 4930 (email:
[email protected])
MPC strategies using state-space, step and impulse re-
sponse linear models with particular application to slow
linear processes have been recorded [1][3]. Despite this
limitation, research activities continued to ourish around
MPC technologies, especially when it was realised that it
can operate plants close to their limits while tackling real-
time constraints. This trend resulted into developing faster
schemes approximated for nonlinear processes, i.e., [4]
[7]. Although this break-through is not devoid of model
approximation errors and computational load occassioned
with slow online optimisation routines found in model-
based schemes which have to be addressed. This paper
proposes a new self-healing algorithm based on model-free
predictive control (SH-MFPC) of DSTATCOM to address
the issue.
Presently, a combination of traditional and intelligent
controls are found in DSTATCOM including variants of
pulse width modulation (PWM) techniques [8]. Their
application became agitated for poor reference following
in the bid to save energy by reducing switching losses
[9]. This drawback currently attracted the adoption of
the ubiquitous proportional-integral (PI) controller and
fuzzy logic controllers, see survey for instance [10]. The
requirement for point-wise online linearization of a non-
gradient algorithm excludes the PI, and fuzzy controllers
on the other hand, from smart control system application
due to heavy rule-base and a lack of credible stability
criteria [11]. In this regard, the novel SH-MFPC scheme
under saturation describing functions of an augmented
nonlinear DSTATCOM model is conditioned to provide
the required control rate at each time instatant in a look-
ahead fashion. The idea is to bypass the model-based
computational burden and predict the presence of limit
cycles for enhancing stability. Integrating a non-gradient
simplex optimiser in the MFPC routine rather than the
usual QP used in [12] guarantees efcient online tuning.
Simulation of the dynamic and steady-state operations
of the new method using SimPowerSystems model has
been evaluated and its performance in solving specic
power quality (PQ) problems i.e., harmonics control com-
pared against conventional optimisers in Matlab/Simulink
environment. The reminder of the paper is organized
as follows. In section 2, analysis of the nonlinear con-
trol switching of the DSTATCOM through the describ-
ing functions (DF) is performed. Section 3 prepares the
needed procedure for augmented DSTATCOM through
the describing functions. An evaluation of the model-free
predictive control methodology and its application to the
distribution system is performed in section 4. Model-free
2
Fig. 1. Describing function: vs. a
optimization problem formulation using simplex routine
is described in section 5. The self-healing property of
the MPC which makes it unique to other controllers was
incorporated and compared within the new MFPC scheme
to further verify its superiority in section 6. Discussions
on simulation results followed in section 7. Experimental
harmonic control application has been explored and anal-
ysed in section 8. While the paper is concluded in section
9.
II. DESCRIBING FUNCTION THEORY
Describing functions theories represent a powerful
mathematical means for evaluating the behaviour of non-
linear systems from [13]. With them, all real systems are
nonlinear except within a limited operating range. Thus,
relieving the control engineer from having to linearise
nonlinear models. This feature is important in tapping
DSTATCOMs nonlinearities to generate desirable peri-
odic signals, and also for determining the stability of
illusive nonlinear feedback control systems e.g., fuzzy
controllers [14]. In off-line MPC design, stability can
easily be achieved through parameter tuning while the
problem is being formulated [4]. However, that is difcult
in on-line MPC design, especially where model accuracy
requirement is paramount. Therefore, our online model-
free design uses the capabilities of DFs to sense the pres-
ence of limit circles in closed-loop. For MFPC, The same
parameters are used to tune the controller for optimality as
in the model-based. The underline assumption being that
all linear syatems are periodically sinusoidal whose period
and amplitude can be predicted by describing functions
depicted in Figure
Figure
It then follows that choosing the error e=a cos t would
simplify extracting the control signal as [14].
u = ue = u(a cos t, a sin t) (1)
Since t has a minimum period of 2 with e as
a trignometric function, then u is a piecewise smooth
function of the same period. Similarly u can be expanded
by considering the fundamental component of a Fourier
Fig. 2. Nonlinear feedback control loop
series such that u = a
1
cos t + b
1
sin t while ignoring
the sub-harmonics and subjecting all the even copontents
to zero, so that
a
1
=
_
2

__

0
u(a cos t, a sin t) cos tdt (2)
b
1
=
_
2

__

0
u(a cos t, a sin t) sin tdt (3)
the describing function, DF may thus be deduced as
N(a, ) =
a
1
jb
1
a
(4)
III. AUGMENTED DSTATCOM
The nonlinear DSTATCOM model can be given in the
ordinary differencial equation as
x = f(t, x, u) (5)
y = g(t, x, u) (6)
where the variables f and g can be evaluated using
any analytical solver, while u and y are the external
control input and output signals to the DSTATCOM.
The augmented describing function of the saturation for
DSTATCOM model based on (4) can then be derived as
N(a, ) = 1, if 1 (7)
N(a, ) =
2

(sin
1
()+(
_
(1
2
))), if 1 < < 1
(8)
N(a, ) = 1, if 1 (9)
For limit cycle to exist [14],
G(jw) = P(jw)D(jw) =
1
N(a)
(10)
where a is the amplitude of the input sinusoidal signal.
Equation (10) is however difcult to solve analytically.
But, it is easy to determine the presence of a limit cycle
that courses instatbility at each intersection of the right
and the left hand terms of the equation using Nichols
curve. The right hand term gives the amplitude of the limit
cycle which depends on the saturation droop, while the
3
Fig. 3. Nichols chart for describing function
left hand term gives its frequency that is independant of
the droop size. The describing function maps between the
lines, irrespective of the droop size at (0dB, -180) and
(, -180) as depicted in Figure
IV. MODEL-FREE PREDICTIVE CONTROL PARADIGM
The concept of the model-free control is built around
the model-on-demand theory described in [3]. The idea is
to perform modelling around unique operating points on
the input-output data by directly storing it in pairs such
that for every modelling requirement around that operating
point, a subset of data is made available from the data base
for modelling operation, instead of indirectly adapting the
global process model. This scheme offers the advantage of
realizing a congruent model free from linearized operating
point dispersions. For simplicity, the three stages adopted
for realising a model-free method from [11], e.g. restrained
control structure, optimization criteria, and optimization
routine, are applied here.
A. Restrained control structure
On-line operations of model in Figure
g(t) = n((t)) +(t), t = 1, . . . N (11)
where g(t) is a nonlinear model updated at every new
operating point, n(.) is the undened models nonlinear
mapping and (t) is a white noise term. A decoupled
predictive controller with a high gain factor and an inte-
grator can be replicated and included in the feedback as a
conventional proportional-integral scheme in the following
way:
P() =
1
+

2
s
(12)
where the controller parameters are represented by
P(s) P(), k
p

1
, k
i

2
, and
i
,
i = 1, . . . , l. While the vector of the parameters is
expressed as
i

l
. An integrator is included at this
stage to enable limit cycle capture and smoothens out
oscillations in steady state as shown in Figure
Fig. 4. Integral effect on limit cycles
B. Gradient Optimization Criterion
LQG cost functions have been used since the 1960s
to solve optimization problems based on gradient infor-
mation in plants with large time-constatnts [11]. But,
for high speed systems of small time-constants where
the operating point is constantly changing, a direct and
robust optimization criterion is required to access an online
database to solve an optimization problem. The simplex
algorthm is a direct search method introduced in the next
section with such properties. The description of a basic
deterministic objective function routine ideal for localising
weighting indices for error, control, and process output
signals expressed in the continuous time domain is given
as
J() =
1
2T
f
_
T
f
0
[R
y
e(t, )
2
+
2
S
u
u(t, )
2
]dt (13)
for simplicity, the wighting lters in (10) are set to R =
S = 1, then the gradient cost function formulation becomes
J() =
1
2T
f
_
T
f
0
(e
2
(t, )) +
2
(u
2
(t, ))dt (14)
this formulation leads to the generation of four control
parameters based on the vector
l
shown as
J

=
1
T
f
_
T
f
0
_
e(t, )
e(t, )

+
2
u(t, )
u(t, )

_
dt
(15)
Steady state is achieved as T
f

C. Non-Gradient Simplex Optimization
The gradient in (12) tends to slow down the routine
and increases compuational count especially in nonlinear
systems. As the papers main contribution, a simplex
algorithm is incorporated as an iterative feedback tuner
(ITF) within the restrained prediction space to search for
optimal control solution of the following minimization
problem.
min
w.r.t.
l
J() (16)
4
The simlex IFT guarantees gradient-free functions that
can practically access online closed-loop data by bypass-
ing the intermediate models often found in model-based
designs. The method is achieved by decomposing the
gradients in (12) to a simple quadratic problem solvable
over nite coincidence horizon C and a control horizon
M in direct online MPC formulation described in [15].
Replacing (13) with a new design objective J :
M

+
to synthesize the MPC to solve for Q :
M
such
that
J(u) = min
u
||e|| (17)
subject to constraints imposed by the describing func-
tion saturation on the actuators and its rates of change
given by
u
min
u(t) u
max
: u
min
u(t) u
max
(18)
and
e = r y (19)
is the error to be minimized, while y is the predicted
output at the current time t estimated from the model
and actual output information. Numerical minimization
is preferred in practice as analytical zeroing of J means
setting innite control gains which threatens the systems
stability [4]. Thus, the new online formulation of (14)
taking (13) into consideration yeilds
J(u) = min
u(t+1),...,u(t+M)
_
L+P1

k=L
|r(t +k) y(t +k)|
r
_
1
r
(20)
the objective becomes clear from (17) of minimizing
the control effort with respect to the error. To realise
this with a given setpoint value s(t), a reference trajectoy
r(t) will be traced by the plant output y(t). Such desired
trajectory is obtained by the response of a rst-order
low-pass pre-lter in [15]. However, low-pass lters have
low input leakage effect in their pole-zero transfer form
which causes poor performance due to attened cross over
frequency. We remedied this effect by replacing the lter
with an inherently stable nite impulse response (FIR)
lter that requires no feedback. This means that all its
poles are located at the origin within the limit cycle,
an essential condition for stability. The impulse response
f(n) is then calculated by feeding the FIR lter input
with the Kronecker delta (n) impulse signal to yeild a
response of a set of coefcients or tap weights b
n
using
f(n) =
N

i=0
b
i
(n 1) = b
n
(21)
for n = 0 to N. The discrete implementation is
realized in Z-plane as
F(z) = Z {f[n]} =

n=
f(n)z
1
=
N

n=0
b
n
z
n
(22)
The reference trajectory r is also dened in the
continous-time domain as
r(t) = W(1 e
t

) (23)
In order to compute J through (17) discrete-time do-
main to conform with (19), a difference equation is derived
from (20) as
r(t + 1) = r(t) + (1 )W (24)
for a chosen control interval T;
= 1
T

(25)
Hence, iterating
r(t +2) = r(t +1) +(1 )W =
2
r(t) +(1
2
)W
(26)
Generally, the reference trajectory is iteratable k-steps
ahead such that
r(t +k) =
k
r(t) + (1
k
)W (27)
Similarly, it is possible to estimate from the plant model
D(t) the output y(t + k) by iterating yM(t) k-steps
ahead. As the objective function is being formulated, its
minimization can be applied online to obtain an optimal
control sequence of the form
u = [u(t + 1)u(t + 2), . . . , u(t +M)] (28)
D. Choice of the Performance Index
To enhance our choice, we went beyond the traditional
ITAE and IAE time domain indices normally found in
CAutoD package and added the derivative of the error ( e)
to the basic index as a requirement for high accuracy and
low chattering in steady state, to construct the performance
criteria shown in Figures
J = J
ITASE+ITASED
=
n

t
t|e
2
(t)| +
n

t
t| e
2
(t)| (29)
where J
ITASE
and J
ITASED
are the cost of the re-
spective model components; e(t) and e(t)) are the same
as above under no-load condition, and n is the nal-time.
The performance as applied to various solvers is compared
in Table 1.
V. CONSTRUCTING SIMPLEX TUNING
As earlier explained, the simplex version of the iterative
feedback tuining introduces the concept of computing a
gradient-free cost function in our model-free design. The
5
TABLE I
COMAPARATIVE PERFORMANCE BETWEEN DIFFERENT SOLVERS
S/No Optimiser Best Solution FEV
1 Pattern Search 0.0046 136
2 Threshold Acceptance 10.9867 1019
3 Simulated Annealing 0.0046 1325
4 Simplex 0.0046 69
5 Simple GA 0.0046 1620
6 EA-CAutoD 0 3031
Fig. 5. Minimisation plot in 3D
template of the algorithm is required for a given control
parameter vector (k)
l
to perform online control
optimization through the gradient extension of (12) as
J

(k) =
1
T
f

T
f
0

e(t, (k))
e(t, (k))

+
2
u(t, (k))
u(t, (k))

dt
(30)
The aim is to compute the closed-loop system signals
at varying operating points. This way, real system data is
made available to the optimal routine at all time through
the database storage offered by the closed-loop signals
namely; reference w(t), error e(t), manipulated variable
u(t) and control variable y(t), prescribed as per Figure
2 in the time domain. again refer to [4] for details on
Fig. 6. Maximisation plot in 3D
Fig. 7. Model-free signal generation
such computations. From the on going, implementing the
model-free routine of (17) clearly requires processing the
actual plant data to obtain all the signals in (15) i.e.,
e(t, (k)),
e(t, (k))

, u(t, (k)),
u(t, (k))

(31)
as depicted in Figure
VI. SELF HEALING PROPERTY OF MFPC VS. MPC
One of the unique and most interesting feature of the
MPC which has been inherited by the MFPC is the look-
ahead property. When the controller is operated respec-
tively using prediction and control horizons of 20 and
2 in both unconstrained MFPC and MPC designs. The
MFPC automatically exhibits this mode and anticipates
faults while they occur and provide predicted response to
mitigate them. Figure
Fig. 8. MFPC in look-ahead mode
VII. SIMULATION RESULTS
The MFPC routine is tested in simlink for a steady
state operation where its performance is compared to other
6
Fig. 9. Control effort for MFPC look-ahead
Fig. 10. MPC in look-ahead mode
Fig. 11. Control effort for MPC look-ahead
Fig. 12. DSTATCOM nonlinear responses to step change
Fig. 13.
MFPC
= 0.0045 rads:
PI,PD,FPD
= 0.0046 rads
designs in closed-loop path. Responses from the PI, PD,
FPD and MFPC shemes presented in Figure
The superiority of the novel method is further veried
in Figure 15 by comparison with the conventional MPC
which exhibited a huge undershoot subsequently followed
by ringing before settling in about 2 cycles. The control
signals producing the effect are respectively shown in
Figure 16 with MPC exerting (0.004528 rads) twice as
much effort required by the new MFPC (0.002934 rads).
Fig. 14. Responses to step change
7
Fig. 15. Step change in alpha
Fig. 16. PI compensated THD = 35.59%
VIII. HARMONICS CONTROL
To demonstrate the level of harmonic cancelation
achieved by the various control strategies investigated
using experimental DSTATCOM simulation model. It is
observed in Figure 17 that the ubiquitous PI controller
offers the worst THD control of 35% for a closed-loop
compensation of the fundamental (220 volts) which was
reduced to 83.74 volts. While the novel MFPC scheme
offers the best control in Figure 18 with the THD re-
duction of 0.08%, almost replicating the fundamental
signal. Conversely, both FPD and PD approaches have
shown promising results over the ubiquitous PI with THD
trends of 12.35% and 21.31%, while reproducing the
fundamental to 141.6 and 139.8 volts, respectively shown
in Figures 7.24c and 7.24d. Further comparison with an
experimentally controlled PCC voltage using three-and-
ve level converters, which realised THD reduction of
1.6% and 0.5% from (Sensarma, 2000) still proves the
superiority of our novel solution by the MFPC. The THD
of 0.08% result achieved is the new record and has by
far satised the 5% IEEE Std. 519 1992 (IEEE, 1993)
requirement.
Fig. 17. MFPC compensated THD = 0.08%
Fig. 18. FPD compensated THD = 12.35%
IX. COCLUSION
In this paper, a novel MFPC technique has been de-
signed and simulated in steady-state for VoltVar optimi-
sation to be used in controlling power quality problems
at a distribution level via DSTATCOM. Its performance
was duly compared and found superior to other schemes
by offering lower control signal of 0.0045 radians against
0.0046 radians from the rest, at a no-load custom voltage
compensation (220 volt feeder). A careful choice of the
performance index lead to a remarkable improvement to
Fig. 19. PD compensated THD = 21.31%
8
the respose speed through a non-gradient simplex solver.
While its performance in controlling harmonics exper-
imentally proved better to other schemes, offering the
record lowest harmonic level ever achieved within the
industry standard IEEE Std. 519 - 1992. The results are
promising for a competing practical application in a smart
distribution grid.
REFERENCES
[1] C. E. Garcia, D. M. Prett, and M. Morari, Model predictive
control: Theory and practice a survey, Automatica, vol. 25, p.
330, 1989.
[2] M. Nikolaou, Model predictive controllers: A critical synthesis
of theory and industrial needs, M. Nikolaou, Model Predictive
Controllers: A Critical Synthesis of Theory and Industrial Needs,
Advances in Chemical Engineering,, vol. v 26, pp. 131204, 2001.
[3] T. Barry and L. Wang, A model-free predictive controller with
laguerre polynomials, in 5th Asian Control Conference, v 1, p
177-184, 2004., vol. 1, 2004, pp. 177184.
[4] J. M. Maciejoweski, Predictive Control with Constraints, J. M.
Maciejoweski, Ed. Prentice Hall, 2002.
[5] A. Stenman, Model-free predictive control, in In Preceedings of
the 38th IEEE Conference on Decision and Control 3712-3717,
Dec 1999, pp. 37123717.
[6] Y. Wang and S. Boyd, Fast model predictive control using online
optimization, in Proceedings of the 17th World Congress the
International Federation of Automatic Control, Seoul, Korea, July
2008, pp. 69746979.
[7] T. Zanma and N. Asano, Off-Line Model Predictive Control
of DC-DC Converter Model Predictive Control www.sciyo.com,
ser. Model Predictive Control www.sciyo.com. 51000 Rijeka,
Croatia: Sciyo Janeza Trdine 9, September 2010, ch. 12, pp.
269282, accessed from www.sciyo.com. [Online]. Available:
www.sciyo.com
[8] A. G. Beccuti, Hybrid control techniques for switching-mode dc-
dc converters part ii: The step-up topology, in Proc. ACC, 2007,
pp. 54645471.
[9] K. J.

Astr om and T. H agglund, Method and an apparatus in tuning
pid regulator, U.S. Patent 4 549 123, 1985.
[10] B. B. Bukata and Y. Li, Reviewing dstatcom for smart distri-
bution grid applications in solving power quality problems, in
Proceedings of the 17th International Conference on Automation
& Computing (ICAC11), University of Hudderseld, Hudderseld,
UK, September 2011.
[11] M. A. Johnson and M. H. Moradi, PID Control, New Identication
and Design Methods, M. A. Johnson, Ed. Springer-Verlag London
Limited, 2005.
[12] S. Hjalmarsson, H; Gunnarsson and M. Gevers, A convenient iter-
ative restricted complexity control design scheme, in Proceedings
Conference on Decision and Control, Lake Buena Vista, Florida,
December 1994, pp. 17351740.
[13] J. C. Hsu and A. U. Meyer, Modern Control Principles and
Applications, J. C. Hsu, Ed. Mc-Graw Hill, 1968.
[14] D. P. Kwok, P. Tam, C. K. Li, and P. Wang, Analysis and design
of fuzzy pid control systems, in IEEE International Conference
on Control, vol. 2. IEEE, August 2002, pp. 955960.
[15] H. Kashiwagi and Y. Li, Nonparametric nonlinear model predictive
control, Korean Journal of Chemical Engineering, vol. 21, no. 2,
pp. 329337, 2004.
[16] CyboSoft. (2010, 07) Mfa vs pid control comparison. [Online].
Available: www.cybosoft.com/newsevents/pidvsmfa.html
[17] C. N. M. Canale, M. Milanese, Semi-active suspension control
using fast model-predictive techniques, IEEE Transactions on
Control System Technology 14, pp. 10341046, 2006., vol. 14, p.
10341046, 2006.