You are on page 1of 13

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO.

6, NOVEMBER 2016 4953

Analysis on Applicability Problems of the


Aggregation-Based Representation of Wind
Farms Considering DFIGs’ LVRT Behaviors
Jia-Yang Ruan, Zong-Xiang Lu, Member, IEEE, Ying Qiao, and Yong Min

Abstract—Applicability problems of the widely-used aggre- can be aggregated as well. The aggregated model contains me-
gation-based representation of DFIG wind farms focus on the chanical, electrical, protective parts and their controllers, and
modelling “error and impact” in transient stability computa- simplifies multiple WTs into one with a rescaled capacity. Obvi-
tion. In order to identify crucial LVRT behaviors of DFIGs,
specifically in system-level analysis, control strategies including
ously, the detailed wind profile or network topology in the wind
vector-orientation, closed-loop rotor-current regulation, and farm (WF) is not required in system-level modelling, and active
Crowbar protection are given a full consideration. It is proved power predicted by wind forecast systems is taken as the input,
that, the DFIG works with its electrical and mechanical dynamics which is quite compatible with system operators.
decoupled, and the result is that the DFIG behaves as an ideal Instead of proposing a novel aggregated model, this paper
controlled-power source. The DFIG's power-modulation functions mainly focuses on the so-called applicability problem, which
defined by the grid code distinguish it from the conventional
synchronous generator, which also conveniently offers dominant
refers to modelling “error and impact” when we use one
dynamics for studying errors of aggregation. Next, errors caused large-capacity equivalent WT to represent small-capacity
by the coupling between distribution factors and LVRT behaviors ones. Although it is common that the WTs have identical
of wind farms (but ignored in aggregation) are discussed. With control strategies and parameters, it's almost unavoidable that
some existing grid codes, these factors not only generate a dis- the aggregated model still well deviate from the detailed model
tributive profile of residual voltages along the feeders, but more in special situations, as will be proved in this paper.
importantly nonuniform power responses among individual wind
turbines. Impacts of aggregation's errors on transient stability of
As for applicability study, “waveform accuracy” is surely a
power systems are analyzed. The trend in the error and impact of main target, but more importantly, system operators are eager
aggregation is analytically surveyed in a rudimentary system, and to know what the rationale of errors is, and what the crucial
further numerically verified in a multi-machine system. impacts are to power systems if errors exist, despite of the fact
that aggregation has already been widely applied in industry.
Index Terms—Aggregation, nonuniformity, transient stability,
low voltage ride-through (LVRT), collection network, wind To evaluate the error and impact, most of current researches
distribution. were carried out in a “scenario-dependent” way. The errors were
verified by comparing time-domain responses of detailed and
aggregated models under a given wind profile [3], [4]. Limita-
I. INTRODUCTION tions of this methodology include: inner-plant distributions of
the detailed wind profile or collection network are required in
these studies, yet they are actually invisible or at least meaning-

I N system-level computation, it is unrealistic and unneces-


sary to integrate all individual DFIG wind turbines (WTs)
into simulation. In fact, large numbers of WTs have to be mod-
less to operators; it is unclear whether the errors can be general-
ized to other wind profiles or other WFs. Some works focused
on improving the waveform accuracy using modified aggrega-
eled by aggregation [1], [2], both in theoretical analysis and tion methods, either aggregated electrical parts of the WF while
industry. The most typical case is to aggregate all WTs in a retaining individual mechanical parts [5], [6], or tried to make
power plant into an equivalent one. In the same way, several WTs represented by a multi-machine model, by utilizing “co-
wind farms connected to the same point of common coupling herency-based” grouping techniques [7], [8]. These researches
did help reducing waveform errors in the studied scenarios, but
numerical results could not illuminate the impacts of aggrega-
Manuscript received August 28, 2015; revised December 07, 2015 and tion on transient stability computation.
February 14, 2016; accepted March 05, 2016. Date of publication March 15,
2016; date of current version October 18, 2016. This work was supported in For more general conclusions about errors of the aggrega-
part by National Key Technologies R&D Program (2013BAA01B03) and in tion, the methodology should be beyond the model representa-
part by National Natural Science Foundation of China (51190101). Paper no. tion itself and be more conceptual: to identify which behaviors
TPWRS-01203-2015.
The authors are with the State Key Laboratory of Power Systems, De- of individual WTs and the WF are crucial for transient stability
partment of Electrical Engineering, Tsinghua University, Beijing 100084, analysis, and to pick out the source of error that is inevitable in
China (e-mail: ruanjiayang@gmail.com; luzongxiang98@tsinghua.edu.cn; aggregation.
qiaoying@tsinghua.edu.cn; minyong@mail.tsinghua.edu.cn).
Color versions of one or more of the figures in this paper are available online
In the power systems, what really matters is the “dominant
at http://ieeexplore.ieee.org. dynamics”, or rather electromechanical behaviors, and all
Digital Object Identifier 10.1109/TPWRS.2016.2539251 other “trivial dynamics” of generators are to be as simplified

0885-8950 © 2016 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.
4954 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 6, NOVEMBER 2016

as possible. Classically, these trivial dynamics include stator


flux transients, PLL (phase-locking loop) and internal fast vari-
ables. LVRT behaviors were considered to be closely related
to both dominant and trivial dynamics of a DFIG [9], [10], but
these two types of dynamics were not explicitly separated for
system-level analysis.
In essence, errors of aggregation result from the complex,
non-linear or even discontinuous dynamics of the converter
interface. During a transient like LVRT, dynamic behaviors of a
DFIG are determined by both protection and control schemes,
and they could hardly be modeled by classical dynamical sys-
tems mathematically. These peculiarities distinguish the DFIG Fig. 1. Controllers of rotor-side excitation in a DFIG.
from the conventional synchronous generator, nevertheless
were not given a sufficient attention in the past.
As a basis for indicating errors of aggregation, DFIGs' dom- differences between a converter-interfaced generator and a
inant dynamics during LVRT are studied in Section II. It is conventional generator.
proved that a DFIG acts as a controlled electrical-power source, When in the doubly-fed state, the open-loop model of induc-
therefore the metric of evaluating its external characteristics is tion generators in the terminal voltage -oriented reference
quite different from the classical “angle-swing behaviors” of the frame is
synchronous generator. In addition, the Crowbar protection is
inspected to know about its effects on DFIGs' controlled-power-
source-like nature. The concept of applicability problems is pro- (1)
posed in Section III. Working points and locations of individual
DFIGs are diverse in a real WF, and the error of aggregation is where the state variable is stator flux and inner EMF
directly generated by this diversity. In Section IV, “nonunifor- (electro-motive force): [ ; input refers
mity-incurred errors” are explicated in details adopting a prac- to the rotor-side excitation voltage and terminal voltage:
tical grid code. Effects of nonuniformity in distribution of the [ . The equation in (1) is flux and voltage
collection network and wind speed are systematically discussed. equations of DFIGs, where the stator current is composed of
Given errors are inevitably introduced by aggregation, the fun- state variables:
damental impacts on transient stability of a rudimentary system
are studied in Section V. In Section VI, a multi-machine system
is adopted to testify the proposed methodology. Applicability (2)
problems of aggregation in different working conditions and
fault conditions are summarized. The output is rotor current ; the output equation
is
II. LVRT BEHAVIORS OF AN INDIVIDUAL DFIG
Compared to the conventional synchronous generator, the
(3)
DFIG wind turbine is interfaced with the grid through a
power-electronic converter, with which the LVRT capability Fig. 1 shows the rotor-side controllers of DFIGs, where cur-
can be realized. The capability not only means that the DFIG rent commands are calculated from the reference power
should be able to keep in operation during a disturbance, but instantaneously:
also the DFIG is required to fast modulate its electrical power
following the predefined rules.
In this section, it will be proved that the vector-oriented con-
trol endows the DFIG with the capability to modulate electrical (4)
powers as an ideal controlled-power source. Quite distinguished
Next, rotor currents are accurately regulated by PI controllers.
from conventional synchronous generators, novel converter-in-
By synthesizing the models, bode diagrams of the rotor cur-
terfaced generators like the DFIG have their dominant dynamics
rent loop can be calculated, shown in Fig. 2(a). The results indi-
(power modulation) decoupled from slow mechanical variables
cate that the controller has a very wide bandwidth, with a cut-off
in system-level analysis. The cumbersome effect of Crowbar
frequency of about 57 Hz. From the step response in Fig. 2(b),
protection is studied, and the switching of Crowbar should be
the rise time of actual currents is no longer than 0.017 s (one
a trivial dynamic.
cycle of 60 Hz), and the overshoot is negligible. In a word, it
can be asserted that the inner current loop responds to and fol-
A. Controlled-Power-Source-Like Nature of the DFIG
lows the command extremely fast. Hence in the electromechan-
Both magnitude and angle of the DFIG's excitation voltage ical time-scale, the actual power is regarded as
can be fast regulated. Comparatively, the excitation voltage tracking the reference almost with no time delay
phasor of a synchronous generator is fixed to the slow me- or overshoot, and DFIGs can be considered to be close to a con-
chanical rotor angle. This explains the critical and essential trolled electrical-power source.
RUAN et al.: ANALYSIS ON APPLICABILITY PROBLEMS OF THE AGGREGATION-BASED REPRESENTATION 4955

so the initial rotor flux can be obtained from and at :

(7)

In the static reference frame, now the stator flux and rotor
flux are determined by zero-input differential equations:

(8)

where the stator impedance includes impedances of


the unit transformer and transmission line at the generator's ter-
minal; is the total leakage factor: .
In (8), and decay at the transient time constant and
, respectively, and . The two
time constants are very low for induction generators, consid-
ering the transmission line connected to the stator has a high
ratio, and Crowbar resistance is high ( is a part
of as it is connected into rotor windings in series). Typical
values are s, s. From the relation be-
Fig. 2. Closed-loop responses of DFIGs' current controllers: (a) the bode dia- tween currents and fluxes, the current is
gram of , (b) the step response of .

B. Crowbar Effects on LVRT Behaviors (9)

As a DFIG is partially coupled to the grid through a con-


verter, Crowbar is installed to avoid overcurrent. Obviously it The peak value is reached approximately at the time
would interrupt DFIGs' controllability on electrical powers. Yet . Replacing the time into (9), we can obtain the peak
since the discontinuous behavior is difficult to accurately sim- value:
ulate using classical electromechanical models, it is intuitively
neglected in many power-system softwares by considering
Crowbar effects as minor. The simplification should be further (10)
interpreted.
The most commonly-used activation strategy is: if reaches which confirms that low values of and would effectively
the cut-in threshold , Crowbar is immediately triggered and suppress the overcurrent in rotor windings.
hence the controller blocks the converter, making the gener- The result is that, the rotor current decays to an acceptable
ator work as an uncontrollable squirrel-cage induction gener- level very quickly after Crowbar is triggered, ideally no longer
ator. As soon as the rotor current decays to an acceptable level than one cycle of 60 Hz. Then the converter can be unblocked
(cut-out threshold), Crowbar is deactivated and the converter is and regain controllability. The peak current is also declined sig-
unblocked in no time. This strategy is robust and reliable, but nificantly when Crowbar resistance is with enough ohms.
the activation time has to be very short, since the DFIG absorbs During a severe fault, unblocking the rotor-side converter
excessive amount of reactive power when rotor wingdings are might again cause overcurrent, and Crowbar is activated the
temporarily short-circuited. second time. In fact, it is not rare that Crowbar is switched on/off
The short-circuit current can be estimated approximately in several times in the early stage of severe faults, as shown in
p.u. systems. The pre-fault initial stator flux is Fig. 3, where the total pulse width of the triggering signal is
only about 15 ms. As the uncontrollable time is very short, the
(5) switching would not (and it should not) have a drastic impact on
modulated powers and thus LVRT behaviors of DFIGs. Fig. 4
After fault occurs, the magnitude of rotor currents increases compares responses of a DFIG installed or not installed with
fast till Crowbar is triggered at . The stator flux basically Crowbar: external characteristics of electrical power are con-
stay unchanged till since . At , the magni- sistent; only little difference emerges in the early stage of faults
tude of rotor currents is equal to the cut-in threshold : due to switching of Crowbar.
For system-level analysis based on electromechanical
(6) models, it is practical to ignore Crowbar effects, namely, a
4956 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 6, NOVEMBER 2016

Fig. 3. Activation and deactivation of Crowbar in the early stage of faults.


Fig. 6. China's grid code for WTs' uninterrupted operation.

TABLE I
TWO COMMON ACTIVE-POWER MODULATION CODES (DURING FAULTS)

above the solid line. Besides the requirement of uninterrupted


operations, the WT is requested to provide a reactive power
Fig. 4. Power responses of a DFIG with/without Crowbar protection. adaptive to the terminal voltage :

(11)

As for the active power , the code only specifies that the
WT should recover to the initial state at a fixed rate of at least
0.10 p.u./s after fault clearing.
During-fault modulation of is not yet specified due to
lack of knowledge on how this behavior impacts the adjacent
power systems, and now it is still intuitively defined by manu-
facturers themselves. Two possibly existing codes are listed in
Table I.
Fig. 5. General illustration of dominant dynamics in LVRT behaviors. Considering the terminal voltage seldom dips to below
0.20 p.u. (unless the fault location is on the 35 kV feeder right
in the WF), and the fault clearing time of modern protection
DFIG is simplified as in the controllable doubly-fed state unless relays is no longer than 300 ms, all WTs are assumed to stay
it is disconnected from the grid permanently. connected to the grid during the voltage dip.
To sum up, the DFIG quickly enters the contingent state when When a number of WTs executing the same grid codes
voltage dips: reducing and lifting up . After fault are clustered, there are two main factors of a detailed WF
clearing, it restores at a fixed ramp rate till it recovers to that cannot be reflected in the aggregated one: nonuniform
the initial state. The wind farm has similar power responses, distribution of the wind speed, and distribution of the collec-
which offers available “dominant dynamics” to evaluate the tion network (35 kV feeders). Both these two factors lead to
error of aggregation. Main concerns of the dynamics include: so-called “nonuniformity” among WTs. The incurred errors
“Remaining Active Power”, “Incremental Reactive Power”, and can be comprehensibly illustrated using the simplest topology,
post-fault “Recovery Time”. See features of these dominant dy- with the number of individual WTs in Fig. 7, which
namics in Fig. 5. is the elementary unit of any general WFs. In the aggregated
model Fig. 7(b), a lumped network is placed to account
III. APPLICABILITY PROBLEMS OF AGGREGATION for the inner-plant losses of the 35 kV feeders.
The applicability problem refers to the “error and impact” As Section II implies, the main concern during faults is the
of aggregation-based representation of wind farms (WFs). The modulated electrical power, which is measured at low-voltage
topic mainly focuses on: though commonly all wind turbines side of the step-up transformer in Fig. 7. In Fig. 7(a), the two
(WTs) in a WF execute the same grid code, when will the re- individual WTs undergo different “terminal residual voltage”
sponse of the aggregated model well deviate from that of the and , whereas the two feeding branches, along with
real WF? the aggregated model, are exposed to the same “grid residual
Fig. 6 is a regular code for WTs (China's national standard). voltage” . In Fig. 7, the centralized control is sometimes
WTs must stay connected to the grid when fault conditions are adopted in application to regulate steady-state power flows. Its
RUAN et al.: ANALYSIS ON APPLICABILITY PROBLEMS OF THE AGGREGATION-BASED REPRESENTATION 4957

where definition of and is shown in Fig. 7(a). Compara-


tively, in the aggregated model, is

(15)

So the error boils down to the relation between (14) and (15),
which is the mathematical definition of “convexity” of the func-
tion . The aggregated model produces more active power
only when is convex, whereas less when is con-
cave. The aggregated model introduces no error if and only if
is globally non-convex.
The conclusions also apply to the reactive power . Er-
rors incurred by nonuniformity of wind distribution can be
explained by convexity of function and as well.
Convexity of these functions is directly decided by grid codes,
and in reality (13) is not always strictly global non-convex func-
tion. For industry, an immediate inference is that, the system op-
erators can minimize errors of aggregation by adjusting the grid
codes.

Fig. 7. The elementary unit of wind farms ( stands for susceptance


IV. CASE STUDY 1: “NONUNIFORMITY-INCURRED
of the feeders): (a) the detailed model and (b) the aggregated model. ERROR” WITH GRID CODE B
In this section, errors incurred by nonuniformity are studied
and verified, where the practical grid code B in Table I is
effect is not a major concern, considering the additional com- adopted.
mands are distributed to each WT uniformly in most cases:
, and the control does not take effect A. Calculation of the Modulated Power in Fault States
during LVRT because WTs only use local measurement when The reactive power of an individual wind turbine (WT)
entering LVRT mode (to modulate following grid codes). is
In Fig. 7, constrained to the aggregation criteria:
(16)

where is the -axis rotor current. The power controller mod-


ulates the through adjusting , in order to satisfy the re-
(12)
quested rule (11). If the limit of current is reached, yet
where superscript stands for the “pre-fault” working point, and is still not adequate, then the rest of requested is
stands for susceptance of the 35 kV feeder, errors of electrical generated by the grid-side converter. Otherwise if there is a sur-
powers are reflected by the errors between and , and plus:
the errors between and .
(17)
As all WTs having identical parameters is the most common
situation, during-fault power in p.u. system is uniquely deter- then the surplus can be used to maintain the active power .
mined by the susceptance of the transmission line , the pre- The stator active power is determined by and :
fault working point , and the residual voltage . In fact, mod-
ulated electrical power of each feeding branch can be expressed (18)
in a general form:
The total active power output by the generator is

(19)
(13)
where is the active power flowing through the grid-side con-
To analyze “nonuniformity-incurred errors” resulting from verter, and it is proportional to the generator slip.
the difference between and in Fig. 7(a), pre-fault working “Remaining Active Power” of an individual WT is directly
points (wind speed) of the two individual WTs are set to be iden- related to the terminal residual voltage , shown in Fig. 8,
tical: . Hence in the detailed model, where the pre-fault working point is p.u. Obvi-
satisfies ously there exist three controlling regions during faults. When
the residual is high, defined as the general type of “Minor
Faults”, the WT can maintain pre-fault . When is low,
defined as the type of “Serious Faults”, the WT has to reduce
(14) to 0, no matter the pre-fault working points. Except for
4958 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 6, NOVEMBER 2016

Fig. 8. “Remaining Active Power” of an individual WT during LVRT (pre-fault


working point is 0.70 p.u.).

Fig. 10. Illustration of errors in “Remaining Active Power” ( p.u.,


Fig. 9. Convexity of “Incremental Reactive Power” as a function of the feeder's p.u.): (a) controlling regions of each WT and (b) comparisons
susceptance . between detailed and aggregated models.

the two extreme situations, the WT enters the “descending re- terminal voltages among WTs, the generated nonuniformity
gion” when the so-called “Ordinary Faults” occur. and incurred errors in electrical power are barely observable.

C. Nonuniformity in the Wind Distribution


B. Nonuniformity in the Collection Network
Wind distribution affects pre-fault working points of
In reality, the length of a single 35 kV feeder seldom exceeds
WTs. To independently observe the active-power nonuni-
50 km, or rather, its reactance is no more than
formity caused by the wind distribution, it is assumed that
(20) .
The basic idea is that, although pre-fault working points of
which is much lower than that of the unit transformer (0.12 p.u.). the two individual WTs might differ a lot, the error can still be
Thus it can be coarsely anticipated that the nonuniformity in low if they reduce the active power uniformly. For example,
susceptance of the collection network is inconspicuous. when code A in Table I is used, active-power reductions of the
Intuitively, the nonuniformity in mainly generates devia- two individual WTs are both equal to %, so the aggregated
tions of the reactive power. Fig. 9 shows the relation between model is still applicable.
and at different levels of the grid residual voltage . Clearly, Nonetheless this is not the case with code B. Assume pre-
though the requested rule (11) is nonlinear, still has low fault working points of the two individual WTs are:
convexity in most of the range. Actually the modulated p.u., p.u., and hence the aggregated model's output
is insensitive to variation of , so the following equation is is p.u. “Remaining Active
satisfied: Power” of each WT is plotted in Fig. 10(a), where axis is
the grid residual voltage . If an “Ordinary Fault” occurs, say
(21) dips to 0.60 p.u., the WT with pre-fault working point 0.30
p.u. still maintains its pre-fault working point:
which means that nonuniformity in the collection network can p.u. Meanwhile, the other individual WT with the pre-fault
be tractably averaged by selecting a proper lumped collection working point 0.70 p.u. enters the “descending region”:
network . The only problem is that, when the feeder is very p.u. Thus of the detailed model is:
long (or is low in Fig. 9), the curve shows some convexity, p.u. Comparatively, since the working point of the aggregated
which incurs errors in the “Incremental Reactive Power”. mode is p.u., it still stays in the “maintaining
The same analysis is carried out for “Remaining Active region”: p.u.
Power” as well, and the conclusions are basically identical: If a more severe fault occurs, say dips to 0.40 p.u., the two
is quite insensitive to variation in . Hence though distribution individual WTs in the detailed model, along with the aggregated
of the collection network and do cause deviations of model, all enter the “descending region”. In this case, output
RUAN et al.: ANALYSIS ON APPLICABILITY PROBLEMS OF THE AGGREGATION-BASED REPRESENTATION 4959

Fig. 11. Convexity of electrical powers as a function of working points : Fig. 12. Errors in recovery behaviors: (a) “Recovery Time” with different pre-
(a) Remaining Active Power and (b) Incremental Reactive Power. fault working points and (b) profile of time-domain comparisons.

of the aggregated model can accurately represent output of the produces less Incremental Reactive Power than the detailed one,
detailed model. This is clearly shown in Fig. 10(a) as well. As a as shown in Fig. 11(b).
result of the nonuniformity among individual WTs, there exists After fault clearing, the two individual WTs in the detailed
an approximate interval of the grid residual voltage , in which model restore active power in the same ramp rate 0.10 p.u./s,
the detailed model produces lower “Remaining Active Power” thus their “Recovery Time” is different, as shown in Fig. 12(a).
than the aggregated one, shown in Fig. 10(b). The nonuniform post-fault behaviors could also incur errors
As stated in Section III, this “nonuniformity-incurred error” of aggregation. In general, the aggregated model recovers ac-
can be rigorously explained by studying convexity of the tive power faster, because recovery of the real detailed WF is
function . Fig. 11(a) shows that, the “Remaining Active obviously restricted to the WT that recovers the most slowly,
Power” (blue solid line) is composed of two separate convexity which should not be explained by convexity-based analysis. In
zones ( p.u. and p.u.), which makes the Fig. 12(a), the WT with p.u. recovers in no time
complete curve a convex function. Errors are generated only after fault clearing, thus in Fig. 12(b) it offers no help in re-
when working points of the two individual WTs span different covery (blue solid line).
convexity zones. When they are in the same convexity zone,
the outputs are superimposable, thus no error exits. From (14)
V. CASE STUDY 2: IMPACTS OF ERRORS
and (15), it is certain that the error must fulfill
EXISTING IN AGGREGATION
(22) The aggregation is likely to incur errors in “Remaining Active
Power”, “Incremental Reactive Power”, and “Recovery Time”
The above case p.u., p.u. is plotted in produced by a wind farm (WF), which would bring about mis-
Fig. 11(a). taken estimation in transient stability.
The nonuniformity in wind distribution would lead to errors A rudimentary system containing a conventional power plant
of reactive power as well, since the error of generates ex- and a WF is surveyed, shown in Fig. 13. A three-phase fault
cessive reactive-power losses in the collection network. “Incre- occurs in the middle of the tie line , with the short-cir-
mental Reactive Power” is the blue solid line in Fig. 11(b). The cuit reactance . Before fault occurs, the active power of the
curve is also composed of two convexity zones, one is convex synchronous generator SG is 1.0 p.u, and the power
( p.u.), and the other is non-convex ( is p.u. Connecting reactances are:
p.u.), yet the complete curve behaves as a concave function. The p.u. ( p.u.), p.u,
figure indicates that errors of “Incremental Reactive Power” are p.u., p.u.
closely related with wind distribution as well, and both From analysis in previous sections, DFIGs behave as an ideal
and can be observed in reality. In the above controlled-power source. Inevitably, generators of this type
case p.u. and p.u, the aggregated model bring in nonlinear network constraints into the system, leading
4960 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 6, NOVEMBER 2016

Fig. 13. A rudimentary system integrated with a DFIG wind farm.

to the so-called “singularity-induced instability” [11], [12].


The mechanism of voltage collapse shall never be observed in
conventional power systems composed of only synchronous
generators.
Dynamic models of the rudimentary system can be expressed
in the form of differential-algebraic equations (DAE), in which
the SG is represented by the classical 2nd order model, along
with the network constraints:

(23)
(24)

where the state variable is the speed deviation and power angle
of the SG: ; the auxiliary variable is voltages
of each node: in Fig. 13; and
can be taken as control commands of WF.
A critical fact is that, the network constraint is nonlinear
if or . Generally, the so-called “singu- Fig. 14. Power-angle curves and determinant of the sensitivity matrix with
larity” of power systems is closely related to the nonlinearity. different power-modulation behaviors: (a) active-power modulation and (b) re-
active-power modulation.
The sensitivity matrix is the partial derivative of to each
node voltage :

(25) Obviously the constraint is kept nonlinear, and hence singularity


is still possible to occur.
According to the famous “implicit function theorem”, as long The integration-termination means that dynamics of the
as is nonsingular, node voltages is directly determined by power system can no longer be modeled by (23) and (24) from
by a nonlinear equation: the instant on. In fact, some faster dynamics would start to take
effect, e.g., stray capacitance effects or the electromagnetic
(26) transient of the grid network, which would lead to instanta-
neous uncontrollable states, e.g., overcurrent or over/under
So the system trajectory can be integrated uniquely and voltage in the local system. The DFIG would immediately lose
smoothly by replacing (26) into (23): controllability of its electrical power and be tripped. Hence
occurrence of singularity is always followed by fast collapse,
(27) defined as “transient voltage collapse”.
In the rudimentary system, and are the only two state
Nonetheless, when the power angle swings after disturbance, variables, making it convenient to analyze the stability mecha-
it is possible that the matrix is approaching singularity nism using classical “power-angle curves”. Power-modulation
. Once singularity occurs, the system voltages can no longer of WF would affect the shape of power-angle curves of SG, see
be solved from (26), and the integration of (27) is immedi- Fig. 14, where axis represents the power angle of SG , and
ately terminated. In essence, the singularity comes from non- axis represents SG's active power.
linearity of the network constraint , which results from the Fig. 14(a) indicates that, when WF temporarily reduces its
controlled-power-source-like nature of DFIGs. In some extreme active power during faults, the power-angle curve of SG
cases, the DFIG has to enter the constant-current mode owing is lifted, and the curve becomes higher if more is cut
to the descended voltage and the rotor-current limit. So the con- during fault. The rotor acceleration of SG is thus suppressed,
straint (24) becomes: which improves transient angle stability. Meanwhile, it is no-
ticeable that there exist peculiar breaking points separating the
(28) power-angle curves into well-defined and undefined sections,
RUAN et al.: ANALYSIS ON APPLICABILITY PROBLEMS OF THE AGGREGATION-BASED REPRESENTATION 4961

which are called “singularity angle ”. When swings to ,


system singularity occurs, which can be told from at
in the figure. Notably, WF's reducing also causes a re-
markable increase in the value of , so as to push the system
away from potential singularity. In Fig. 14(b), it is shown that
the modulation of reactive power has much less impact on ac-
celeration of adjacent SGs. Its only great impact is improving
the singularity angle.
The results confirm how WF interacts with the synchronous
system through its LVRT responses. When WF is represented
by aggregated models, errors in the simulated power modula-
tion behaviors could have great impacts on estimation of angle-
swing of adjacent SGs and thus CCT of the system.
The anticipations are directly verified in the rudimentary
system in Fig. 13, where WF is still composed of two individual
WTs, with pre-fault working points 1.0 and 0.20 p.u, respec-
tively. When the clearing time is 200 ms (guaranteeing both
cases are stable), simulation results are plotted in Fig. 15. There
do exist errors in “Remaining Active Power” and “Incremental
Reactive Power” of WF, and the biases are consistent with
the analysis above. Impacts are that, SG maintains a lower
breaking torque in Fig. 15(c), leading to a gain of the peak
power-angle in Fig. 15(d). Note that aggregation also brings
about errors in “Recovery Time” in Fig. 15(a), that is, the
“Recovery Time” becomes shorter. Impacts of WF's faster
recovery would limit SG's post-fault capability to decelerate,
also shown in Fig. 15(c), (d).
The calculated critical clearing time (CCT) is a vital index
for operators. The simulation result is: the CCT is 253 ms if WF
is modeled by detailed models, whereas the index deteriorates
to 225 ms when WF is modeled by aggregated models. The
instability is clarified in Fig. 16. After fault clearing, the angle
swing of SG in Fig. 16(a) continues to increase till about 170
degrees, at which singularity occurs, signified by the sudden
divergence of WF's voltage in Fig. 16(b). In reality, WF would
trip at about s, when the collapsed voltage dips to an
unacceptable level. The abrupt change of voltage waveforms is
quite distinguished from classical angle stability problems. The
results also show that, if voltage collapse happens, WFs are still
to be tripped even if they have reliable LVRT capabilities.
Fig. 16(c) is rotor currents of the aggregated WF. When the
fault is cleared, the rotor current starts to gradually increase as
the active power is in the ramp recovery mode as designed. At
s, the rotor current has reached the maximum limit
owing to the low voltage in Fig. 16(b). At s, the
rotor current is getting un-controllable (along with the collapsed
voltage), also indicating that singularity-induced instability oc-
curs, although the WF is not in constant-power mode at this
instant.

VI. CASE STUDY 3: ANALYSIS AND VERIFICATION OF Fig. 15. Simulation results of the rudimentary system with detailed/aggregated
models of WF (fault clearing time is 200 ms; both cases are stable): (a) active
APPLICABILITY PROBLEMS IN A MULTI-MACHINE SYSTEM power of WF, (b) reactive power of WF, (c) active power of SG, and (d) trajec-
tory of SG's power-angle swing.
A real wind farm (WF) composed of 31 wind turbines (WTs)
with is taken for simulation and analysis, as shown
in Fig. 17. For a given level of wind power, nonuniformity-in- distributive errors, corresponding to countless uncertain inner
curred errors arise in the aggregated model. Considering the un- states of a WF. A high level of “error distribution” signifies that
certainty of wind profile, “error distribution” is proposed for in- the contribution of a WF to the power system is getting unpre-
dicating errors of “Remaining Active Power” and “Incremental dictable and non-unique even if its total output power is fixed
Reactive Power”. Given a fixed total power of a WF, there exist to a known value.
4962 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 6, NOVEMBER 2016

Fig. 18. Simulation results of “Minor Faults”: (a) Remaining Active Power and
(b) Incremental Reactive Power.

Fig. 19. Simulation results of “Ordinary Faults”: (a) Remaining Active Power
and (b) Incremental Reactive Power.

Fig. 20. Simulation results of “Serious Faults”: (a) Remaining Active Power
and (b) Incremental Reactive Power.
Fig. 16. Instability of the rudimentary system with aggregated models of WF
(fault clearing time is 225 ms; the system loses stability; the aggregated WF
is tripped at about s): (a) trajectory of SG's power-angle swing,
(b) residual voltage of WF, and (c) rotor current of WF.
The wind farm in Fig. 17 contains two 35 kV feeders, con-
necting 20 and 11 WTs respectively. At each time slice, the de-
tailed wind profile is input into the detailed model (actually in-
accessible for operators), whereas the total wind power is input
into the aggregated model. It is inferred from Section IV that, the
deviation of during-fault power is related with the grid residual
voltage . Therefore at each time slice, three occasions of faults
are applied: Minor Faults, Ordinary Faults, and Serious Faults,
in which the voltage dips to 0.80 p.u., 0.50 p.u, and 0.30 p.u.,
respectively. The results are shown in Figs. 18–20, where “error
distribution” is well illustrated, showing inherent fluctuations
and uncertainties of the detailed WF.
Some common features can be observed in all three fault
conditions. WF is able to maintain its active power when at
low-power pre-fault working points. At high-power working
points, “Remaining Active Power” is reduced owing to limits
of rotor currents. Notably, reactive power also shows distribu-
Fig. 17. Aggregation-based representation of a real wind farm. tive features due to the coupling effects indicated in Section IV:
RUAN et al.: ANALYSIS ON APPLICABILITY PROBLEMS OF THE AGGREGATION-BASED REPRESENTATION 4963

Fig. 21. Illustration of errors existing in aggregation in different pre-fault


working points using the convexity-based analysis.

Fig. 22. IEEE 14-bus multi-machine system for dynamic verification.


Fig. 23. Dynamic responses of the detailed/aggregated WF (fault clearing time
is 200 ms; both cases are stable):, (b) reactive power of WF1, and (c) grid
additional reactive-power losses of the collection network in- residual voltage of WF1.
curred by the errors of “Remaining Active Power”. Errors of
active and reactive power arise in the identical range of pre-fault
working points. Fig. 21, there is quite a probability that the distribution spans
Furthermore, the detailed and aggregated models exhibit little different convexity zones even if WF is in the low-power range,
error when WF is at either low-power or high-power pre-fault and errors would also grow large.
working points, which can be explained using the convexity- In Figs. 18–20, “Remaining Active Power” of the detailed
based analysis in Fig. 21. When pre-fault active power of WF model is generally lower than that of the aggregated model.
is 0.20 p.u., it's reasonable to assume that wind profiles of indi- In Fig. 18(b), “Incremental Reactive Power” of the aggregated
vidual WTs satisfy a distribution law PDF1, with its mean value model is higher than that of the detailed model at medium-power
of 0.20 p.u. Obviously, if the uncertainty of wind distribution pre-fault working points, whereas lower at high-power pre-fault
is low, i.e., the band of PDF1 is narrow, almost all WTs are in working points. The results are consistent with the theoretical
the same convexity zone, as the red solid line shows. There- analysis in Section IV.
fore their “Remaining Active Powers” are superimposable in To sum up, applicability of aggregation modelling deterio-
the low-power range, and so the aggregated model generates rates evidently only in the circumstance when WF is at medium-
little error. This is the case with high-power states as well. Nev- power working points before the fault occurs, especially when
ertheless, it becomes probable that working points of WTs span the system undergoes so-called “Ordinary Fault” disturbances.
different convexity zones in the medium-power range, as the IEEE 14-bus system in Fig. 22 is adopted for dynamic sim-
green solid line shows. Thus the output turns out well deviated ulation. In the multi-machine system, the reference generator
and reveals a wide band of “error distribution” in Fig. 19. Note- SG0 is installed with a speed governor and automatic voltage
worthily, when uncertainty in wind profile arises, e.g., PDF2 in regulation (AVR) controllers, representing a large remote
4964 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 6, NOVEMBER 2016

Fig. 24. Synchronous generator SG1's angle swing when detailed/aggregated


WFs are modeled (fault clearing time is 200 ms; both cases are stable):
(a) during- and post-fault swing and (b) details during the fault.

system. SG1–SG3 are synchronous generators; WF1–WF4


represent four WFs, each of which contains 31 WTs shown
in Fig. 17. In the local system, the installed capacity of wind
energy is comparable with that of conventional power plant,
so the error of the aggregated WF would have great impacts
on transient stability. As inspired by Fig. 19, pre-fault active
power of the four WFs is set to a medium level (0.42 p.u.), and
an “Ordinary Fault” is simulated in order to better observe the
errors. In Fig. 22, a three-phase fault occurs on the tie line L15 Fig. 25. Instability signified by voltage collapse of the system and tripping of
WFs (fault clearing time is 249 ms, and only the case with aggregated WFs is
with the fault impedance 0.001 p.u., making sure that residual unstable): (a) grid residual voltage of WF1, (b) active power, and (c) reactive
voltages of the local system decrease to around 0.45 p.u., so power of WF1.
that the case is an “Ordinary Fault”, as shown in Fig. 23(c). The
fault is cleared by circuit breakers at both terminals of L15.
Initially the fault clearing time is set to 200 ms, making the fault. The result is that, reactive power of aggregated WFs in
two cases with detailed and aggregated models of WFs both Fig. 23(b) becomes higher in a short period after fault clearing,
stable. The responses of WFs are plotted in Fig. 23. As for the since the grid code requests the WFs to provide supporting
applicability, it is clear that during the fault, “Remaining Active “Incremental Reactive Power” if the voltage is lower than 0.90
Power” of the aggregated WF is higher than that of the detailed p.u. The difference in post-fault modulation of reactive power
WF, whereas “Incremental Reactive Power” of the aggregated is also good evidence that SGs accelerate faster when WFs are
WF is lower, which validates the analytical anticipation. represented by aggregation.
In addition, post-fault recovery of WFs' active power shows The impacts of these errors are that, the estimated CCT of the
significant errors. Since the grid code regulates that all WTs system with aggregated WFs is deviated from the real value. If
restore their active power in a fixed ramp rate, it is well reflected the fault clearing time is increased to 249 ms, the real system
that WTs in the power plant do recover to pre-fault working can still maintain transient stability, whereas the system with
points at different instants. Fig. 23(a) confirms that active power aggregated WFs is already unstable. The instability can be sig-
of the aggregated WF does recover faster than the detailed WF. nified by voltage collapse of the system in Fig. 25, where the
As a consequence of the deviation in “Remaining Active grid residual voltage of WFs cannot restore after fault clearing.
Power” of WFs, active powers of SGs are more suppressed In Fig. 25(a), the residual voltage of WF1 descends to an ab-
when WFs are modeled by aggregation. Thus rotors of SGs normal value at about s after fault clearing, followed
naturally accelerate faster than the real situation, as shown by an abrupt divergence, which is an explicit sign of occurrence
in Fig. 24. The error can also be confirmed in Fig. 23, where of singularity. Simultaneously, WFs lose controllability on elec-
the system voltages are lower with aggregated WFs during trical power in Fig. 25(b), (c), followed by immediate tripping
RUAN et al.: ANALYSIS ON APPLICABILITY PROBLEMS OF THE AGGREGATION-BASED REPRESENTATION 4965

of these power plants. The consequence of voltage collapse is [7] M. Ali, I. S. Ilie, J. V. Milanovic, and G. Chicco, “Wind farm model
consistent with the simple results in Fig. 16 observed in the rudi- aggregation using probabilistic clustering,” IEEE. Trans. Power Syst.,
vol. 28, no. 1, pp. 309–316, Feb. 2013.
mentary system. [8] L. Lin, J. Tan, Y. Chen, and W. Y. Liu, “Coherency-based dynamic
equivalent for power system centralized large scale wind power,” in
Proc. IEEE. Int. Conf. Power System Technol. (POWERCON), 2012,
VII. CONCLUSION pp. 11–16.
[9] W. T. Liu, Y. K. Wu, C. Y. Lee, and C. R. Chen, “Effect of low voltage
Applicability of the aggregated model of DFIG wind farms ride-through technologies on the first Taiwan offshore wind farm plan-
ning,” IEEE. Trans. Sustain. Energy, vol. 2, no. 1, pp. 78–86, Jan. 2011.
is given a systematic study. The contribution of the paper in- [10] N. R. Ullah, T. Thiringer, and D. Karlsson, “Voltage and transient sta-
cludes: dominant dynamics of the DFIG are clarified to make bility support by wind farms complying with the E.ON Netz grid code,”
studying the applicability problem possible; “error and impact” IEEE. Trans. Power Syst., vol. 22, no. 4, pp. 1647–1656, Nov. 2007.
[11] J. Ruan, Z. Lu, Y. Qiao, and Y. Min, “Transient stability of wind turbine
of wind farms' LVRT behaviors are proposed for transient sta- adopting a generic model of DFIG and singularity-induced instability
bility problems; the mechanism that wind distribution raises the of generators/units with power-electronic interface,,” IEEE. Trans. En-
error through coupling with improper grid codes is explained. ergy Convers., vol. 30, no. 3, pp. 1069–1080, May 2015.
[12] V. Venkatasubramanian, H. S. Chattler, and J. Zaborszky, “Dynamics
During a voltage dip, the DFIG modulates its electrical of large constrained nonlinear systems—A taxonomy theory,” Proc.
power in a distinctly short time-scale and high accuracy, which IEEE, vol. 83, no. 11, pp. 1530–1561, Nov. 1995.
is essentially different from power-angle behaviors of the syn-
chronous generator. Due to the controlled-power-source-like
nature, three typical power-modulation quantities step up as
the main concern of a wind farm's LVRT behaviors in transient Jia-Yang Ruan was born in AnHui, China, in 1987.
stability. Errors in simulated power responses of a wind farm He received B.S. and M.S. degrees in electrical en-
have direct impacts on voltage stability and acceleration of gineering from Tsinghua University, Beijing, China,
in 2010 and 2012, respectively. His research interests
adjacent conventional generators. include induction motor control, stability of grid-con-
The concept “nonuniformity” among individual wind nected wind turbines, and application of aggregated
turbines is presented as the source of errors. Nonuniformity-in- modelling for wind farms. Currently, he is pursuing
the Ph.D. degree of electrical engineering at Tsinghua
curred errors are shown to be closely related with convexity of University.
the grid code. Distribution of the collection network generates
a nonuniform profile of residual voltages, whereas can be
tractably averaged by selecting a proper lumped network. Dis-
tribution of the wind makes DFIGs with diverse working points
enter different convexity zones in a nonuniform way. Thus the Zong-Xiang Lu (M’02) was born in ChongQing,
China, in 1974. He received B.S. and Ph.D. degrees
wind turbines' electrical powers become non-superimposable, in electrical engineering from Tsinghua University,
and hence errors are produced in aggregation. Last, all DFIGs Beijing, China, in 1998 and 2002, respectively. He is
do not recover to initial working points at the same instant after now an associate professor of electrical engineering
at Tsinghua University, Beijing China, where he has
fault clearing. been employed since 2002. His research interests
In simulation, the anticipated errors and impacts are explicitly include power system reliability, large scale wind/PV
observed. Due to uncertainty of wind profiles, “error distribu- stations integration analysis and control, DG and
microgrid, energy and electricity strategy planning.
tion” arises especially when wind farms are at medium-power
pre-fault working points, which means their impacts on tran-
sient stability is getting unpredictable. The index well reflects
the proposed applicability problems when aggregation is ap- Ying Qiao was born in JiangSu, China, in 1981.
plied to wind farms. She received the B.S. and Ph.D. degrees in electrical
engineering from Shanghai JiaoTong University,
Shanghai, and Tsinghua University, Beijing, China,
in 2002 and 2008, respectively. She is now a lecturer
REFERENCES of electrical engineering at Tsinghua University,
Beijing, China, where she has been employed since
[1] V. J. Marandi, L. F. Pak, and V. Dinavahi, “Real-time simulation of
2010. Her research interests include renewable
grid-connected wind farms using physical aggregation,” IEEE. Trans.
energy and power system security and control.
Ind. Electron., vol. 57, no. 9, pp. 3010–3021, Sep. 2010.
[2] V. Akhmatov, “Analysis of dynamic behaviour of electrical power sys-
tems with large amount of wind power,” Ph.D. dissertation, Dept. Elect.
Power Eng., Tech. Univ. Denmark, Kgs. Lyngby, Denmark, 2003.
[3] J. Brochu, C. Larose, and R. Gagnon, “Validation of single- and
multiple-machine equivalents for modeling wind power plants,” IEEE.
Trans. Power Convers., vol. 26, no. 2, pp. 532–541, Jun. 2011.
[4] A. Shafiu1, O. A. Lara, G. Bathurst, and N. Jenkins, “Aggregated wind Yong Min was born in HuBei, China, in 1963. He
turbine models for power system dynamic studies,” Wind Eng., vol. 30, received the B.S. and Ph.D. degrees in electrical en-
no. 3, pp. 171–186, 2006. gineering from Tsinghua University, Beijing, China,
[5] M. A. Chowdhury, W. X. Shen, N. Hosseinzadeh, and H. R. Pota, in 1984 and 1990, respectively. He is currently a pro-
“A novel aggregated DFIG wind farm model using mechanical torque fessor with the Department of Electrical Engineering,
compensating factor,” Energy Convers. Manage., vol. 67, no. 1, pp. Tsinghua University, Beijing, China. His research in-
265–274, Mar. 2013. terests include power system stability and control.
[6] L. M. Fernandez, C. A. Garcia, and J. R. Saenz, “Aggregated dynamic Prof. Min is a Fellow of the IET.
model for wind farms with doubly fed induction generator wind tur-
bines,” Renew. Energy, vol. 33, pp. 129–140, 2008.

You might also like