You are on page 1of 10

CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING, VOL. 42, NO.

3, SUMMER 2019 163

Measurement-Based Characteristic Curves for


Voltage Stability and Control at the Point of
Interconnection of Wind Power Plants
Courbes caractéristiques basées sur les mesures
pour la stabilité et le contrôle de la tension au
point d’interconnexion des centrales éoliennes
Xiaodong Liang , Senior Member, IEEE, Md. Nasmus Sakib Khan Shabbir, Student Member, IEEE,
Nahidul Khan, and Xiaodi Yan
Abstract— In this paper, the measurement-based characteristic curves, reactive power–voltage ( Q–V ) curve
and reactive power–active power–voltage ( Q– P–V ) curve, are developed for voltage stability and control at
the point of interconnection (POI) between a wind power plant (WPP) and a utility grid. The procedure to
obtain an effective Q–V curve using the supervisory control and data acquisition (SCADA) measurement data
is proposed. The SCADA data used in this paper were recorded in 2016 at the POI of a 27-MW WPP currently
in operation in Newfoundland and Labrador (NL), Canada. Each Q–V curve is created for a particular month
using the SCADA data corresponding to a very narrowly defined active power range, so the active power is
considered constant as the conventional Q–V curve concept. The MATLAB curve fitting toolbox is used to
realize the most fit mathematical equations for Q–V and Q– P–V curves. The developed Q–V and Q– P–V
curves can be used for the data-driven voltage control for WPPs; and the Q–V curves can also be used for
the voltage stability evaluation.
Résumé— Dans cet article, les courbes caractéristiques basées sur la mesure, la courbe puissance réactive-
tension (Q–T) et la courbe puissance réactive-puissance active-tension (Q–P–T) sont développées pour la stabilité
de la tension et le contrôle au point d’interconnexion (PI) entre une centrale éolienne (CE) et un réseau de
distribution. La procédure permettant d’obtenir une courbe Q–T efficace, à l’aide des données de mesure du
système de contrôle et d’acquisition de données (SCADA), est proposée. Les données SCADA utilisées dans ce
document ont été enregistrées en 2016 au PI d’un CE de 27 MW actuellement en opération à Terre-Neuve et
Labrador (TNL), Canada. Chaque courbe Q–T est créée pour un mois donné à l’aide des données SCADA
correspondant à une plage de puissance active très étroitement définie. La puissance active est donc considérée
comme constante comme dans le concept classique de courbe Q–T. L’outil d’ajustement de courbe MATLAB
est utilisé pour réaliser les équations mathématiques les mieux ajustées pour les courbes Q–T et Q–P–T. Les
courbes Q–T et Q–P–T développées peuvent être utilisées pour le contrôle de tension piloté par les données
pour les CE; Les courbes Q–T peuvent également être utilisées pour l’évaluation de la stabilité de la tension.
Index Terms— Measurement-based reactive power–voltage ( Q–V ) curve, reactive power–active power–
voltage ( Q– P–V ) curve, voltage control, voltage stability, wind farms.

I. I NTRODUCTION To evaluate voltage stability, various indices can be obtained


through characteristic curves. The most commonly used char-
V OLTAGE stability and control are major power system
operation concerns, which become especially challeng-
ing when power grids are populated with large-scale renew-
acteristic curves are real power–voltage (P–V ) and reactive
power–voltage (Q–V ) curves to determine the static voltage
able power generation, such as wind power plants (WPPs). stability limits of a power system [1]–[12]. Several other
characteristic curves have also been reported in the litera-
Manuscript received October 31, 2018; revised January 1, 2019; accepted ture for voltage stability assessment, including the complex
March 11, 2019. Date of current version July 23, 2019. This work was
supported in part by RDC Ignite R&D fund under Grant 5404-2059-101. power–voltage (S–V ) curve [10] and voltage–current (V –I )
(Corresponding author: Xiaodong Liang.) curve to estimate the critical load value at the verge of
X. Liang, M. N. S. K. Shabbir, and X. Yan are with the Department of voltage collapse, which can be obtained using local measure-
Electrical and Computer Engineering, Memorial University of Newfoundland,
St. John’s, NL A1B 3X5, Canada (e-mail: xliang@mun.ca; mnskshabbir@ ments of voltage and current magnitude without requiring
mun.ca; xy0516@mun.ca). the knowledge of the Thevenin equivalent parameters of a
N. Khan is with the Energy Control Center, Newfoundland and Labrador bus of interest [12], the active power–reactive power (P–Q)
Hydro, St. John’s, NL A1B 4K7, Canada (e-mail: nahidulkhan@nlh.nl.ca).
Associate Editor managing this paper’s review: Om Malik. curve to determine the maximal reactive margin at load
Digital Object Identifier 10.1109/CJECE.2019.2906007 buses based on the bus impedance [13], [14], and the active
0840-8688 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
164 CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING, VOL. 42, NO. 3, SUMMER 2019

power–reactive power–voltage (P–Q–V ) curve to assess the


voltage stability margins using the Thevenin equivalent circuit
parameters; it shows the boundary of the maximum active
power, maximum reactive power, and minimum voltage in
a 3-D diagram [15], [16]. In these indices, the voltage and
current are magnitudes. Fig. 1. Simplified power system.
Among different types of characteristic curves, the Q–V
curve has been widely accepted by utility companies, such as II. T HEORETICAL C HARACTERISTIC E QUATIONS
Bonneville Power Administration (BPA) and Western Electric- A complex power system can be represented by an equiva-
ity Coordinating Council (WECC), due to its less convergence lent generation and a load in Fig. 1. The active power P and
problem and direct relationship with reactive power compensa- reactive power Q supplied to the load at the load bus can be
tion [1]. A measurement-based Q–V curve is proposed in [1]; expressed as follows [18]:
however, it requires to adjust certain system parameters and EV
to employ a special compensation method in the measurement P = sin(δ1 − δ2 ) (1)
X
procedure to obtain the Q–V curve, which restricts its practical 2
EV V
usage in power system operation. Q= cos(δ1 − δ2 ) − (2)
Primary reactive power reserves in a system include gen- X X
erators operating in voltage control mode, shunt capacitors, where E and δ1 are the magnitude and phase angle of the
and static VAR compensators. When a system does not have generator terminal voltage, V and δ2 are the magnitude and
adequate reactive reserves or transmission lines fail to deliver phase angle of the voltage at the load bus, and X is the
the required reactive power to the location in need, voltage magnitude of the branch reactance between the generation and
instability occurs. Real-time corrective actions are required, the load.
such as adding shunt capacitors as proposed in [17]. To achieve Rearranging (1) and (2), the following formulas are derived:
optimal control of the added shunt capacitors for the reactive PX
power compensation, a function of reactive power versus real sin(δ1 − δ2 ) = (3)
EV  
power and voltage magnitude (Q–P–V ) is utilized in [17]; V2 X
however, to obtain this function, the manufacturer’s charac- cos (δ1 − δ2 ) = Q+ . (4)
X EV
teristic curves of wind turbine generators are required, which
might not be readily available. Combining (3) and (4), the following relationship is derived:
In this paper, a measurement-based approach for creating    2
PX 2 QX+V 2
Q–V and Q–P–V curves at the point of interconnection (POI) + = 1. (5)
EV EV
of a WPP is proposed through curve and surface fitting
techniques using the supervisory control and data acquisi- Based on (5), the mathematical relationship of the active
tion (SCADA) field measurement data for a practical WPP. power, reactive power, and voltage magnitude at the load bus
There are no special requirements, such as adjusting system can be expressed as follows:
parameters or requests of wind turbine generator characteristic 1
from the manufacturer, in order to develop the curves. The P =± (E V )2 − (Q X + V 2 )2
X
proposed data-driven approach is based on the measurement P is a real number → (E V )2 − (Q X + V 2 )2 ≥ 0 (6)
data during the system’s normal operation. One advantage of 1 1
the measurement-based characteristic curve is not to rely on Q = − V2 ± (E V )2 − (P X)2
X X
the power-flow simulation, and thus, the results are not affected Q is a real number → (E V )2 − (P X)2 ≥ 0. (7)
by the accuracy of the computer simulation model. Field
measurements reflect the properties of individual components There is a plus or minus sign associated with (6) and (7);
in real life; the generated mathematical functions and curves however, only the positive sign is chosen in each equation.
can be directly used in voltage stability assessment of the POI Therefore, the derived final mathematical functions are
bus and/or implemented in the voltage control actions. 1
This paper is arranged as follows. In Section II, mathe- P = (E V )2 − (Q X + V 2 )2
X
matical derivation based on the fundamental power system (E V )2 ≥ (Q X + V 2 )2 (8)
theory is developed to show the relationship between the 1 2 1
active power, reactive power, and bus voltage magnitude. Q=− V + (E V )2 − (P X)2
X X
In Section III, the measurement-based Q–V curve approach
(E V )2 ≥ (P X)2 . (9)
is proposed using the SCADA data of a WPP currently
in operation in Newfoundland and Labrador (NL), Canada, However, to determine P–V , Q–V , and Q–P–V functions
and the procedure for creating such a curve for a general from (8) and (9), E and X values must be known for a specific
application is proposed. In Section IV, the measurement- system. These parameters can be difficult to determine for a
based Q–P–V curve approach is proposed using the large, complex power grid.
SCADA data for the same WPP. Conclusions are drawn On the other hand, the measurement-based P–V , Q–V , and
in Section V. Q–P–V curves can be constructed through curve fitting, thus
LIANG et al.: MEASUREMENT-BASED CHARACTERISTIC CURVES FOR VOLTAGE STABILITY AND CONTROL AT THE POI OF WPPs 165

curves tool [19]. A Q–V curve has a base case with reactive
power equal to zero. This base case is also named an initial
voltage point.
For a selected bus, if the specified voltage can be reached
without needing additional reactive power compensation when
varying the voltage setting points of the bus, the Q–V curve
for this particular bus may be flat. One example would be to
study the Q–V curve at a real generator’s terminal bus; the
Q–V curve will be flat at zero within the generator’s reactive
power capability. This remains true if the measurements are
taken to obtain the Q–V curve instead of simulation.
Although Q–V curves created through power-flow simula-
tion have been practiced for decades, the accuracy of such
Q–V curves depends on the accuracy of the power-flow sim-
ulation model. To overcome this limitation, a measurement-
based Q–V curve method is proposed in this paper. By using
this measurement-based method, a power system simulation
model and the related procedures are not required; Q–V curves
Fig. 2. Electrical single-line diagram of the sample WPP. are created directly using the field measurement data that
reflect real system operation conditions.
offering a convenient way to evaluate the voltage stability and In this paper, measurements were taken at the POI of the
implement the voltage control at the POI of a WPP. WPP. As a generation system, the whole WPP can be treated
as an equivalent generator, and thus, the POI is practically a
III. S YSTEM U NDER S TUDY generator bus.
A 27-MW grid-connected WPP currently in operation in
NL, Canada, is used as a sample WPP in this paper. There A. Proposed Procedure to Draw Q–V Curves
are nine doubly fed induction generators (DFIG), each rated Three scenarios are considered for creating a measurement-
at 3 MW, in this WPP. The WPP is integrated into the power based Q–V curve at the POI of the sample WPP using the
grid’s 66-kV bus through a step-up transformer. The electrical SCADA data: 1) Scenario 1—base case (the initial voltage
single-line diagram of a wind farm is shown in Fig. 2. point) is the lowest voltage point; 2) Scenario 2—base case
The measurements were conducted on the primary 66-kV (the initial voltage point) is the data point with the reactive
side of the 25-MVA step-up transformer through the SCADA power equal to 0 (Q = 0) or near zero; and 3) directly
system for a whole year in 2016. The data sampling rate plotting the Q–V curve using the measurement data. Similar
is 1 sample point/s; each sample point was stamped with to traditional Q–V curves, a fixed active power value is chosen
a corresponding date and time when it was recorded. Each for each Q–V curve.
measurement point consists of three measured data: the voltage The procedure to draw a Q–V curve for Scenario 1 is shown
magnitude in V, the active power in MW, and the reactive in the following.
power in MVAR. 1) For a given month, choose an active power P value from
It was found that the active power supplied by the WPP the SCADA data, for example, P = 4.1–4.2 MW and
significantly varies during the operation every month, ranging then select all data points within this specified active
from 2 to 27 MW. power range.
2) Rearrange the data points based on their voltages
IV. M EASUREMENT-BASED Q–V C URVE increasing from small to large for Scenario 1 with the
Q–V curves are commonly used for voltage stability eval- lowest voltage point as the very first point in the data
uation and voltage control. The traditional way to generate set.
a Q–V curve at a load bus of interest in a power system 3) The voltage of the first data point is named V 10 , and the
is through off-line successive power-flow simulation. The reactive power of the first data point is named Q10 . This
procedure is to first change the selected load bus to a PV bus point is treated as an initial condition, also known as the
in the power-flow simulation and then, record the required “base case;” therefore, it has reactive power equal to 0
reactive power by placing a variable reactive power source, on the Q–V curve. The first point on the Q–V curve is
such as a fictitious generator (synchronous condenser) at the then determined by Q0 = Q10 − Q10 = 0, V = V 10 ,
selected bus for different bus voltage settings (the active power i.e., (Q0, V 10 ).
remains the same for a particular Q–V curve). The voltage 4) The voltage of the second data point is named V 11
set point of this fictitious generator is varied, and its reactive (V 11 > V 10 ), and the reactive power of the second
power output is allowed to be any value needed to meet data point is named as Q11 . The change in the reactive
the voltage set point; this is how the commercial software, power output for the second data point voltage can be
PowerWorld, computes the Q–V curves in the software’s Q–V calculated as Q1 = Q11 − Q10 . The second point
166 CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING, VOL. 42, NO. 3, SUMMER 2019

reactive power, Q10 , was reduced from each measurement


point. The lowest voltage point has a zero reactive power
after this processing. It moves the Q–V curve downward
(if Q10 is positive) or upward (if Q10 is negative) at this
reactive power amount. For P = 4.1–4.2 MW, the lowest
voltage point has (P, Q, V ) = (4.1040 MW, 0.9172 MVAR,
66.9633 kV/1.0146 p.u.); in this case, Q10 = 0.9172 MVAR,
and the Q–V curve moves downward at the amount of
0.9172 MVAR. For P = 26.1–26.2 MW, the lowest volt-
age point has (P, Q, V ) = (26.1998 MW, 2.5662 MVAR,
67.2868 kV/1.0195 p.u.), Q10 = 2.5662 MVAR, and the Q–V
curve moves downward at the amount of 2.5662 MVAR.
The procedure to draw a Q–V curve for Scenario 2 is given
in the following.
1) Similar to Step 1 in Scenario 1, for a given month
of SCADA data, choose an active power P value, for
example, P = 4.1–4.2 MW, and then, select all data
points within this specified active power range
2) Rearrange data points based on their reactive power
increasing from small to large in order to find the
reactive power point equal to or near zero for Scenario 2.
The base case is chosen as a point with the reactive
power equal to zero or near zero if such a point can be
found, or in the case, all reactive power are much larger
than 0; the smallest reactive power point can serve as a
base case. In either case, the reactive power at the base
case is named Q10 .
3) Once the base case is found, the reactive power value of
every single point in the data set will be subtracted by
Q10 . For the sample point i (i =0−n), Qi = Q1i −
Q10 , V = V 1i . The Q–V curve can be obtained using
all the points (Qi, V 1i ) through curve fitting.
In Scenario 2, if a base case with the exact zero reactive
power value can be found, the calculation can be greatly
simplified because the measured reactive power Q(new) at a
new voltage level (other than the initial voltage corresponding
to the base case) will be equal to Q, Q = Q(new) − 0.
In this case, we do not need to calculate Q but directly
use the measured Q to plot the Q–V curve. Although a base
case with an exact zero reactive power is desired, it may
Fig. 3. Q–V curve at P = 4.1–4.2 MW fit by Gaussian 2 model. be very difficult to find such an operating point in real life,
(a) Scenario 1. (b) Scenario 2. (c) Scenario 3. so near zero is also acceptable; in this case, if it is just a near
zero reactive power value, Q for each new voltage levels
on the Q–V curve is then determined by Q1 = must be calculated as follows Q = Q(new) − Qnearzero.
Q11 − Q10 , V = V 11 , i.e., (Q1, V 11 ). Although the value of Qnearzero is small, it still has a value;
5) Similarly, the voltage of the third data point is named therefore, Q calculation cannot be omitted. In reality, for
V12 , and the reactive power of the third data point is a given active power P value, the reactive power may or
named Q12 . Note that V 10 and Q10 remain the same as may not have a near zero point. For example, for P = 4.1–
steps 3 and 4 and all repeating steps. The change in the 4.2-MW range, the reactive power near zero point is found
reactive power output for the third data point voltage can at 0.00203 MVAR, and the base case is (P, Q, V ) =
be calculated Q2 = Q12 − Q10 . The third point on the (4.189 MW, 0.00203 MVAR, 67.8249 kV/1.0276 p.u.). In this
Q–V curve is then determined by Q2 = Q12 − Q10 , case, the reactive power at every data point is subtracted by
V = V 12 , i.e., (Q2, V 12 ). 0.00203 MVAR. However, for P = 26.1–26.2-MW range,
6) This process will continue until reaching the last data the reactive power near zero point does not actually exist,
point. The new processed data are used to obtain the the one closest to zero is found at 2.4028 MVAR, the base case
Q–V curve through the curve fitting. is (P, Q, V ) = (26.1081 MW, 2.4028 MVAR, 1.0201 p.u.),
In Scenario 1, the reactive power value at the lowest and the reactive power at every data point is subtracted
voltage point serves as a reference; the same amount of by 2.4028 MVAR.
LIANG et al.: MEASUREMENT-BASED CHARACTERISTIC CURVES FOR VOLTAGE STABILITY AND CONTROL AT THE POI OF WPPs 167

TABLE I
G OODNESS OF F IT S TATISTICS FOR P OLYNOMIAL 5 AND G AUSSIAN 2 F ITTING

Scenario 3 is a curve fitting effort directly using the mea- TABLE II


surement data without processing, which reflects the actual G OODNESS OF F IT S TATISTICS FOR P OLY 11 AND
operating range of the WPP. By creating the curve like this, P OLY 22 FOR Q– P –V C URVE
it indicates that there is an operating point at the POI, which
can serve as the base case with the reactive power equal to
zero; it may be out of the narrow operating range of the
measurements, but it should exist within the system.
Fig. 3 shows the curve fitting results of the constructed Q–V
curves using the proposed procedures for P = 4.1–4.2 MW
for the three scenarios. model) and Gaussian 2 (second-order Gaussian model) are
Due to ever-changing operating conditions of the WPP, considered as the target fitting functions for the Q–V curve.
which is significantly different from the conventional power The fifth-order polynomial model can be expressed by the
generation, it is hard to determine a fixed operating point. following equation:
By comparing the three scenarios for a Q–V curve, it is
y = a1 x 5 + a2 x 4 + a3 x 3 + a4 x 2 + a5 x + a6 (10)
found each scenario associated with a Q–V curve in a similar
shape but shifted from each other due to different reference where a1 –a6 are the parameters to be fit. The second-order
points taken in each scenario. Considering that it is reasonable Gaussian model can be written as follows:
to assume that a zero reactive power operating point will (x−b1 )2 (x−b2 )2
− −
always exist although it may not be displayed in the narrow c12 c22
y = a1 e + a2 e (11)
field measurement data, Scenario 3, by directly creating a
Q–V curve using the SCADA data through curve fitting, where a1 , a2 , b1 , b2 , and c1 , c2 are parameters to be fit. In (10)
is recommended to be a simple way to represent a Q–V curve and (11), x is the voltage magnitude in per unit (p.u.) and y
at the POI of a WPP. is the reactive power in MVAR.
To construct a family of Q–V curves, the SCADA
data in March 2016 are processed; the data are assorted
B. Mathematical Equations for Q–V curves based on 13 different real power values: P = 2.1–2.11 MW,
To create a Q–V characteristic curve and obtain the math- 4.1–4.11 MW, 6.1–6.11 MW, 8.1–8.11 MW, 10.1–10.11 MW,
ematical equations, the curve fitting technique is implemented 12.1–12.11 MW, 14.1–14.11 MW, 16.1–16.11 MW, 18.1–
in this paper using the MATLAB curve fitting toolbox. 18.11 MW, 20.1–20.11 MW, 22.1–22.11 MW, 24.1–
As investigated in [2], Polynomial 5 (fifth-order polynomial 24.11 MW, and 26.1 – 26.11 MW. A small range of the active
168 CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING, VOL. 42, NO. 3, SUMMER 2019

TABLE III
E RRORS B ETWEEN C ALCULATED D ATA U SING P OLY 11 AND P OLY 22 IN M ETHOD 1 V ERSUS SCADA M EASUREMENT D ATA IN M ARCH 2016

power is taken instead of a single value in order to obtain in Tables I and II and the corresponding reactive power and
enough points for a better curve fitting while maintaining a voltage values measured in March 2016. Method 2 is to use all
practically constant real power P value for each Q–V curve. SCADA data in March 2016. The two methods are discussed
The goodness of fit statistics of the curve fitting using the in the following.
two target equations, (10) and (11), is displayed in Table I.
The goodness of fit for the fitting equations can be evaluated A. Method 1
by four parameters: the sum of squares due to error (SSE),
In this method, only a partial of the SCADA data in
the root-mean-squared error (RMSE), the R-square, and the
March 2016 is used. Thirteen active power P values shown
adjusted R-square. An accurate model usually has a smaller
in Tables I and II, together with all associated reactive power Q
SSE or RMSE value but with R-square and adjusted R-square
and voltage V values, serve as input data for the surface fitting.
closer to 1 [2].
Both polynomial 11 (Poly11) and polynomial 22 (Poly22) are
It is found that Gaussian 2 model shows better accuracy for
considered as the target functions. The Q–P–V curve in the
most cases than Polynomial 5 model for March data, showing
form of Poly11 is expressed as follows:
smaller SSE and RMSE errors and larger R-square values. The
determined coefficients for Gaussian 2 mathematical models Q 11 = f 11 (P, V ) = p00 + p01 P + p10 V (12)
for each Q–V curve with a particular active power P are
where p00 , p01 , and p10 are coefficients for Poly11, P is the
provided in Table VIII. Three sample Q–V curves using
active power in MW, and V is the voltage magnitude in p.u.
Gaussian 2 for three different active power values are shown
The Q–P–V curve in the form of Poly22 is expressed by
in Fig. 4.
the following equation:
A further study using the SCADA data measured in
  
June 2016 for the same WPP indicates that Polynomial 5 Q 22 = f 22 (P, V ) = p00 + p01 P + p10 V
model has better accuracy for more than half cases for the   
+ p02 P + p11
2
V P + p20 V2 (13)
June data set. Therefore, either Polynomial 5 or Gaussian 2
where p00  , p  , p  , p  , p  , and p  are coefficients for
model can be chosen to develop a Q–V curve based on the 01 10 20 11 02
particular data set. Poly22, P is the active power in MW, and V is the voltage
magnitude in p.u.
The goodness of fit statistics of the surface fitting using the
V. M EASUREMENT-BASED Q– P–V C URVE two target equations, (12) and (13), are displayed in Table II.
The Q–P–V curve is required to implement a control action It appears that Poly 11 and Poly 22 have similar accuracy in
for the voltage at the POI. Reference [17] derived such a this fitting, with Poly 22 slightly more accurate than Ploy 11.
relationship based on the manufacturer data of wind turbines. Coefficients for Poly 11 and Poly 22 with 95% confidence
In this paper, a measurement-based Q–P–V characteristic can bounds can be found in Table IX.
be determined through the surface fitting technique using the The accuracy of the fit Poly11 and Poly22 equations are
SCADA data, and a practical control scheme can be developed compared with the SCADA data in March 2016 by choosing
using this curve. the random measurement data points throughout the month as
In this paper, two methods are considered to develop the shown in Table III; relative errors between the calculated data
Q–P–V curves using the SCADA data for a particular month. using the fit equations and the SCADA measurements are also
Method 1 is to use only partial measurement data, for example, shown in Table III. Table V also indicates that Poly11 has a
use data points sorted with 13 specific active power values comparable error as Poly22.
LIANG et al.: MEASUREMENT-BASED CHARACTERISTIC CURVES FOR VOLTAGE STABILITY AND CONTROL AT THE POI OF WPPs 169

Fig. 4. Created Q–V curves by Gaussian 2 model using the SCADA data at the POI of the WPP measured in March 2016. (a) P = 2.1–2.11 MW.
(b) P = 12.1–12.11 MW. (c) P = 20.1–20.11 MW.

TABLE IV TABLE V
G OODNESS OF F IT D ATA C OMPARISON B ETWEEN P OLY 21 AND G OODNESS OF F IT D ATA OF P OLY 21 M ODELS
P OLY 55 M ODELS

Because the actual data set is significantly scattered,


the robustness algorithm is adopted to improve the accuracy
B. Method 2 of the model. Three types of Poly21 models are studied; one
In this section, the Q–P–V mathematical function is devel- model is without any robustness algorithm, and two models
oped using all SCADA measurement data in March 2016. are with a robustness algorithm. The robustness algorithms
Poly11-to-Poly55 surface fitting models are studied; it is found implemented in this paper are Bisquare and the least absolute
that all the models fit well with the field measurement data set, residuals (LARs) algorithm.
but the accuracy of the model does not significantly improve The goodness of fit results for the three types of
after the Poly21 model. The goodness of fit comparison Poly21 models is shown in Table V. In Table V, the good-
between Poly21 and Poly55 model is tabulated in Table IV. ness of fit data show better fitting for “LAR” and “without
Therefore, to keep the order of the mathematical model robustness” models; the two models have slightly smaller SSE
minimum, Poly21 is considered as the suitable mathematical and RMSE values and slightly larger R-square and adjusted
model. The general expression of Poly21 model is as follows: R-square values than the “Bisquare” model. Coefficients for
the three types of Poly21 models with 95% confidence bounds
Q(x, y) = p00 + p10 x + p01 y + p20 x 2 + p11 x y (14) can be found in Table X.
To further validate the most suitable model, the relative
where p00, p01, p10 , p20 , and p11 are coefficients for Poly21, percentage errors between the calculated reactive power Q
Q (x, y) is the reactive power capability of the WPP in p.u., and the SCADA data using the three types of Poly21 models
x is the voltage magnitude V at the POI in p.u., and y is the are provided in Table VI. Among the five arbitrary chosen
active power P in p.u. The voltage base and power base are operating points in Table VI, the “Bisquare” model shows
66 kV and 27 MVA, respectively. much larger errors than the other two types of models.
170 CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING, VOL. 42, NO. 3, SUMMER 2019

TABLE VI
E RRORS B ETWEEN THE C ALCULATION AND F IELD M EASUREMENT A CTUAL D ATA FOR T HREE T YPES OF P OLY 21 M ODELS

TABLE VII
E RRORS B ETWEEN C ALCULATED D ATA U SING P OLY 21 (LAR) AND P OLY 21 (W ITHOUT ROBUSTNESS ) IN M ETHOD 2 V ERSUS SCADA
M EASUREMENT D ATA IN M ARCH 2016

Fig. 5. Errors between the calculated reactive power Q and SCADA Fig. 6. Developed poly21 model through surface fitting using the LAR
measurement data in March 2016 for models developed by Methods 1 and 2. algorithm based on the field measurement data at POI. (a) Surface fitting
diagram. (b) Residuals.

The “without robustness” model has smaller errors for two


operating points, and the “LAR” model has smaller errors measurement date used in Table III is used to calculate the
for the remaining three operating points. The “LAR” algo- errors between calculations and measurements, and the results
rithm leads to reasonably low and consistent errors, whereas are shown in Table VII.
the “without robustness” algorithm leads to errors that can Tables III and VII show a few large errors, which is rea-
significantly vary. sonable due to the scattered nature of the data; however, most
To further validate the “LAR” and “without robustness” errors are rather small. The four different models developed
models obtained using Method 2, the same set of SCADA using Methods 1 and 2 are compared in Fig. 5 for the errors
LIANG et al.: MEASUREMENT-BASED CHARACTERISTIC CURVES FOR VOLTAGE STABILITY AND CONTROL AT THE POI OF WPPs 171

TABLE VIII
C OEFFICIENTS FOR Q–V C URVES U SING G AUSSIAN 2 M ODEL W ITH 95% C ONFIDENCE B OUNDS

TABLE IX TABLE X
C OEFFICIENTS FOR P OLY 11 AND P OLY 22 (M ETHOD 1) C OEFFICIENTS OF P OLY 21 M ODELS (M ETHOD 2)
W ITH 95% C ONFIDENCE B OUNDS W ITH 95% C ONFIDENCE B OUNDS

between the calculated reactive power Q versus the SCADA


measurement data. It is found that Poly21 (LAR) shows the
smallest and most stable errors among all models. Therefore,
the Poly21 (LAR) model is chosen as the final model in the accuracy of such Q–V curves depends largely on the
this paper. The created model can be readily used in voltage accuracy of the power system simulation model, which is hard
controller design. The surface fitting graph and residuals of to validate in real life, especially when wind power generation
the final model are depicted in Fig. 6. is involved. In this paper, a measurement-based Q–V curve
method is proposed by using curve fitting techniques through
the MATLAB curve fitting toolbox and the SCADA data at
VI. C ONCLUSION
the POI. The Q–V curve is represented by a fitting equa-
In this paper, methods to develop measurement-based char- tion. A fifth-order polynomial equation (Polynomial 5) and a
acteristic curves, Q–V and Q–P–V curves, at the POI of a second-order Gaussian model (Gaussian 2) are investigated in
WPP are proposed. These methods are further implemented case studies as the fitting equation. The procedure of creating
using the SCADA field measurement data recorded at the POI measurement-based Q–V curves is provided in this paper.
of a 27-MW WPP currently in operation in NL, Canada. The Q–P–V curves can be used for the voltage control
Q–V curves are commonly used in power system stability of WPPs; two methods to create the Q–P–V curves at the
analysis or voltage control at a chosen bus. Conventional POI of a WPP using the SCADA data are demonstrated and
Q–V curves are created through off-line successive power- compared. The equations representing the Q–P–V curves
flow simulation for a chosen bus in a power system; however, are determined using surface fitting techniques through the
172 CANADIAN JOURNAL OF ELECTRICAL AND COMPUTER ENGINEERING, VOL. 42, NO. 3, SUMMER 2019

MATLAB curve fitting toolbox. The accuracy of the Q–P–V Xiaodong Liang (M’06–SM’09) was born in
curves is validated using the SCADA measurement data. Lingyuan, China. She received the B.Eng. and
M.Eng. degrees from Shenyang Polytechnic Univer-
The proposed measurement-based methods do not reply sity, Shenyang, China, in 1992 and 1995, respec-
on the power-flow simulation model and do not need the tively, the M.Sc. degree from the University of
manufacturer data or the system configuration. This data- Saskatchewan, Saskatoon, SK, Canada, in 2004, and
the Ph.D. degree from the University of Alberta,
driven approach can help to achieve voltage stability evalu- Edmonton, AB, Canada, in 2013, all in electrical
ation or real-time voltage control with great accuracy. engineering.
From 1995 to 1999, she was a Lecturer with
VII. A PPENDIX Northeastern University, Shenyang. In 2001, she
joined Schlumberger, Edmonton, a world-leading oil field service company,
See Tables VIII–X. where she was promoted to Principal Power Systems Engineer in 2009.
After serving Schlumberger for almost 12 years, she joined Washington State
University, Vancouver, WA, USA, in 2013. From 2013 to 2015, she was an
R EFERENCES Assistant Professor with Washington State University. In 2015, she joined the
[1] Z. Huang, L. Bao, and W. Xu, “A method to measure QV curves and Memorial University of Newfoundland, St. John’s, NL, Canada, where she is
its applications in power systems,” Electr. Power Energy Syst., vol. 29, currently an Associate Professor. Her current research interests include power
pp. 147–154, Feb. 2007. systems, renewable energy, and electric machines.
[2] C. Andalib-Bin-Karim, X. Liang, N. Khan, and H. Zhang, “Determine Dr. Liang is also a registered Professional Engineer in the province of
Q–V characteristics of grid-connected wind farms for voltage control Newfoundland and Labrador, Canada.
using a data-driven analytics approach,” IEEE Trans. Ind. Appl., vol. 53,
no. 5, pp. 4162–4175, Sep./Oct. 2017. Md. Nasmus Sakib Khan Shabbir (S’17) was born
[3] S.-C. Wang, C.-J. Wu, and P.-H. Huang, “Analysis of power system in Tangail, Bangladesh. He received the B.Sc. degree
QV curve by Fuzzy modeling,” in Proc. ICROS-SICE Int. Joint Conf., in electrical and electronic engineering from the
Aug. 2009, pp. 343–347. Rajshahi University of Engineering & Technology,
[4] T. J. Overbye, I. Dobson, and C. L. DeMarco, “Q-V curve interpretations Rajshahi, Bangladesh, in 2012. He is currently pur-
of energy measures for voltage security,” IEEE Trans. Power Syst., vol. 9, suing the Ph.D. degree with the Memorial University
no. 1, pp. 331–340, Feb. 1994. of Newfoundland, St. John’s, NL, Canada.
[5] R. S. Moura, A. C. Z. de Souza, B. I. L. Lopes, and F. W. Mohn, “Effects From 2013 to 2016, he was a Lecturer with the
of QV curves in the dynamic behaviour of power systems,” IET Gener., Department of Electrical and Electronic Engineer-
Transmiss. Distrib., vol. 10, no. 12, pp. 2861–2870, 2016. ing, Bangladesh University of Business and Tech-
[6] B. H. Chowdhury and C. W. Taylor, “Voltage stability analysis: V-Q nology, Dhaka, Bangladesh, where he was promoted
power flow simulation versus dynamic simulation,” IEEE Trans. Power to Assistant Professor in 2016 and served until 2017. His current research
Syst., vol. 15, no. 4, pp. 1354–1359, Nov. 2000. interests include the design and control of high-performance power electronic
[7] P. Kundur, Power System Stability and Control. New York, NY, USA: converters and their applications in renewable energy integration to power
McGraw-Hill, 1994, p. 965. grids.
[8] G. G. Rao, H. P. Kecincha, and D. Thukaram, “Voltage stability: Some
reflections on Q-V characteristics of loads,” in Proc. Int. Conf. Energy
Manage. Power Del. (EMPD), Mar. 1998, pp. 79–84.
[9] Y. Fu, T. S. Chug, and X. Y. Li, “An improved approach to voltage sta- Nahidul Khan was born in Dhaka, Bangladesh. He
bility analysis via network equivalence,” in Proc. Int. Conf. Adv. Power received the B.Sc. degree in electrical and electronic
Syst. Control, Operation Manage. (APSCOM), Nov. 1997, pp. 231–235. engineering from the Khulna University of Engineer-
[10] Y.-Y. Hong and C.-C. Liu, “A heuristic and algorithmic approach to ing & Technology (KUET), Khulna, Bangladesh,
VAR planning,” IEEE Trans. Power Syst., vol. 7, no. 2, pp. 505–512, in 2004, and the M.Eng. and Ph.D. degrees in elec-
May 1992. trical engineering from the Memorial University of
[11] R. R. Austria, X. Xu, and M. Power, “Voltage stability assessment of the Newfoundland, St. John’s, NL, Canada, in 2008 and
National Grid system using modern analytical tools,” in Proc. IEEE/PES 2015, respectively.
Transmiss. Distrib. Conf. Expo., Nov. 2001, pp. 229–234. From 2005 to 2006, he was a Lecturer with Stam-
[12] M. H. Haque, “Use of V-I characteristic as a tool to assess the static ford University Bangladesh, Dhaka. From 2013 to
voltage stability limit of a power system,” IEE Proc.-Gener., Transmiss. 2015, he was a Protection and Control Engineer
Distrib., vol. 151, no. 1, pp. 1–7, Jan. 2004. with Newfoundland and Labrador Hydro - a Nalcor Energy company, St.
[13] M. Sobierajski, K. Wilkosz, J. Bertsch, and M. Fulczyk, “Using bus John’s, where he is currently a System Operation Engineer. His current
impedance and bus P-Q curve for voltage stability control,” in Proc. research and teaching interests include renewable energy integration, energy
IEEE/PES Transmiss. Distrib. Conf. Expo., Nov. 2001, pp. 79–84. storage application, multiterminal high-voltage direct current transmission,
[14] M. Sobierajski and M. Fulczyk, “Voltage stability study by P-Q curve and power system operation and control, developing transient analysis and
with rectangular probability distribution of bus load,” in Proc. 8th Int. creating operational philosophies.
Conf. Probabilistic Methods Appl. Power Syst., 2004, pp. 894–899. Dr. Khan is also a registered Professional Engineer in the province of
[15] S. H. Li and H. D. Chiang, “Continuation power flow with nonlinear Newfoundland and Labrador, Canada.
power injection variations: A piecewise linear approximation,” IEEE
Trans. Power Syst., vol. 23, no. 4, pp. 1637–1643, Nov. 2008.
[16] P. Prabhakar and A. Kumar, “Performance evaluation of voltage stability Xiaodi Yan was born in Beijing, China. He is
index to assess steady state voltage collapse,” in Proc. 6th IEEE Power currently pursuing the bachelor’s degree in electrical
India Int. Conf. (PIICON), Dec. 2014, pp. 1–6. engineering and computer science with the Memo-
[17] W. Edmondson and S. Mukherjee, “Adaptive capacitor switching for rial University of Newfoundland, St. John’s, NL,
wind energy generation,” in Proc. IEEE Ind. Appl. Soc. Annu. Meeting, Canada.
Oct. 2017, pp. 1–6. He has done a co-op work term working on data
[18] Y. Wang, “Advanced load shedding scheme for voltage collapse pre- mining of grid-connected wind-farm measurement
vention,” Ph.D. dissertation, Dept. Elect. Comput. Eng., Univ. Alberta, data at the Department of Electrical and Computer
Edmonton, AB, Canada, 2011. Engineering, Memorial University of Newfoundland,
[19] (Jan. 1, 2019). QV Curves. [Online]. Available: https://www.powerworld. from 2017 to 2017.
com/WebHelp/Content/MainDocumentation_HTML/QV_Curves.htm

You might also like