SET-POINTS RECONFIGURATION IN

NETWORKED DYNAMICAL SYSTEMS

Alessandro Casavola ∗ Domenico Famularo ∗∗

Giuseppe Franzè ∗ Michela Sorbara ∗

∗
DEIS - Università degli Studi della Calabria,
Rende(CS), 87036, ITALY,
{casavola,franze,sorbara}@deis.unical.it
∗∗
DIMET - Università degli Studi di Reggio Calabria,
Reggio Calabria, Via Graziella, 89100, ITALY,
famularo@ing.unirc.it

Abstract: In this paper we present a discrete-time predictive coordination strategy

for load/frequency control problems in networked multi-area electrical power
systems. The aim here is at finding a coordination strategy for the on-line
modification of the prescribed set-points of each power generation unit, so as to
ensure viable evolutions to the overall networked system, with respect to prescribed
operative and safety constraints and despite of load variations and possible faults.
Copyright
c 2006 IFAC

Keywords: Electric power systems, Electrical networks, Set-point control,

Command Governor, Control with constraints.

1. INTRODUCTION control (LFC) (Andersson, 2004), (Bhowmik et

Traditional power system control approaches are
facing several new challenges under deregulation.
Some market-based functions call for greater de-
centralization, as in the allowance of bilateral
contracts (Nobile et al., 2001) and in the Re-
gional Transmission Organizations (RTOs). The
necessity of an open communication infrastruc-
ture is nowadays mandatory for independent gen-
eration companies to offer low-cost third party
services. An improved communication infrastruc-
ture is also of paramount relevance for Indepen-
dent System Operator (ISO), as they are respon-
sible for the monitoring of all network compo-
nents in their jurisdiction. An important aspect
of the system operations is the load/frequency
1 This work has been supported by MURST Project Fault
detection and diagnosis, supervision and control and set-
points reconfiguration.
al., 2004), (Çimen, 1998), (Lim et al., 1997)-
(Wang et al., 1993). The LFC controller is in
charge of keeping the system frequency and the
inter-area exchanged powers as near as possible
to their scheduled values. In fact, an unbalance
in the generated or consumed power will lead to
frequency deviation that, if too large, will have
serious impacts on the system operations. An ex-
ample of such a situation occurs when a gener-
ating unit is tripped due to a failure. In such a
case, effective LFC strategies would be desirable in
order to keep all system variables of interest inside
acceptable limits during transients for avoiding
out-of-services (Andersson, 2004).
Recent papers have addressed this problem from
several points of view. In (Wang et al., 1993), a
robust LFC controller which ensures good per-
formance in the presence of generation rate con-
straints for a single area has been proposed. A

132
two/four-area power system has been considered on the overall network evolutions. Specifically, at
in (Yang et al., 2002) and a solution suggested
via a decentralized control scheme. Saturation has
been included in the turbine model in order to g 1 θ1 y1
Slave
impose generation rate constraints. g 1. θ1
. x1
r. 1 . c1
Here a method based on predictive control ideas, .
g N θN .
. Master .
used recently to synthesize Command Governor rN
x. 1 .
.
(CG) (Bemporad et al., 1997), (Casavola et al., . g N θN yN
xN Slave
2006), (Casavola et al., 2000), (Gilbert and Tin
xN cN
Tan, 1991) and Parameter Governor (PG) units
(Kolmanovsky and Sun, 2006) in more traditional
contexts, is proposed which enforces pointwise-
in-time constraints on the evolutions of relevant Fig. 1. A spatial network of dynamic systems
system variables.
time t, the direct application of the reference
A control strategy is proposed which enforces
r(t) = [r1 (t) . . . rN (t)]T could lead to constraints
pointwise-in-time constraints on the evolutions of
violation. The ROG unit is in charge to mod-
relevant system variables. It consists of adding to
ify such a reference into its best feasible version
a primal compensated system a nonlinear device
g(t) = [g1 (t) . . . gN (t)]T and, if necessary, to add
called Reference-Offset Governor (ROG) whose
a control offset θ(t) = [θ1 (t) . . . θN (t)]T on the
action is based on the actual reference, current
nominal control law.
state and prescribed constraints. The aim of the
ROG device is to modify, whenever necessary, the The paper is organized as follows. In Section 2,
reference and to add an offset on the nominal the ROG scheme is discussed and its relevant
control law in such a way that the constraints properties summarized. In Section 3, a two-area
are enforced and the primal compensated system power system model is described and the problem
maintains its linear behavior. The ROG action is formulated. Computer simulations are finally pre-
computed on-line by solving, at each time session, sented in Section 4 and some conclusions end the
a constrained quadratic programming problem paper.
that usually requires low computational times also
for systems of high order.
In this paper, for the first time upon our’s best 2. REFERENCE-OFFSET GOVERNOR
knowledge, an application of the ROG approach (ROG) DESIGN
to the LFC problem is presented where bounds
on the maximum power and frequency deviations A ROG control scheme, with plant, primal con-
are enforced pointwise-in-time. We focus on a troller (equipped with an integral action) and
two-area power system and the aim is to check ROG device, is depicted in Fig 2. Consider the
the capability of ROG units to reconfigure the
θ(t)
prescribed frequency and/or control offset set-
+ y(t)
points, whenever critical events occur, such as r(t) g(t) Primal u(t)
ROG Controller + Plant c(t)
failures/faults or large unexpected load changes.
Consider the abstract scenerio depicted in Fig. 1,
where the master represents the centralized LFC xc(t) xp(t)
supervisor and the slaves the generation units.
x(t)
There, a single master station is in charge of su-
pervising and coordinating several slave systems,
Fig. 2. The ROG control scheme
which are assumed to be locally compensated dy-
namical systems equipped with independent sen- following linear, time-invariant system of the plant
sors, actuators and (semi)-autonomous decision regulated by the primal controller
capability. In particular, ri , zi , xi , yi and ci rep- 
resent, respectively: the reference, the command  x(t + 1) = Φx(t) + Gg g(t) + Gθ θ(t) + Gd d(t)
y(t) = Hy x(t)
(which includes both a modified reference and a 
control offset), the state, performance-related and c(t) = Hc x(t) + Lg g(t) + Lθ θ(t) + Ld d(t)
coordination-related outputs for the i-th slave sys- (1)
tem. In such a context, the supervisory task can with x(t) ∈ Rn the state vector (which includes
be expressed as the requirement of satisfying some the controller states); g(t) ∈ Rm the manipulable
tracking performance, whereas the coordination reference which, if no constraints were present,
task consists of enforcing for all times t pointwise- it would essentially coincide with the reference
in-time constraints of the form ci (t) ∈ Ci on each r(t) ∈ Rm ; θ(t) ∈ Rm an adjustable offset on the
slave system and/or f (c1 (t), c2 (t), ...., cN (t)) ∈ C nominal control law which we assume be selected
from a given convex and compact set Θ, with

133
0m ∈ int Θ; d(t) ∈ Rnd an exogenous bounded
disturbance satisfying d(t) ∈ D, ∀t ∈ Z+ with D a
specified convex and compact set such that 0nd ∈
D; y(t) ∈ Rm the output, viz. a performance
related signal; c(t) ∈ Rnc the constraints vector
c(t) ∈ C, ∀t ∈ Z+ (2)
which C ⊂ Rnc a prescribed constrained set. It is
assumed that

 1) Φ is a stable matrix;
(A1) 2) System (1) is offset-free w.r.t. g(t) i.e.
 definitions and assumptions, it follows that Wδ
Hy (In − Φ)−1 Gg = Im
In the sequel, let us consider the following state
description of (1)

 x(t + 1) = Φx(t) + Gz(t) + Gd d(t)
y(t) = Hy x(t) (3)
 moreover, it is required that the offset on the con-
c(t) = Hc x(t) + Lz(t) + Ld d(t)
 T
where z(t) = g T (t) θT (t) ∈ R2m is the ROG
output
 and the following
 matrices are defined G =
Gg Gθ , L = Lg Lθ . The ROG design problem
consists of generating, at each time instant t, the
command input z(t) as a static function of the
current state x(t) and reference r(t)
z(t) := z(x(t), r(t)) (4)
in such a way that, under suitable conditions, the
constraints (2) are fulfilled for all possible distur-
bance sequences d(t) ∈ D and possibly y(t) ≈ r(t).
Moreover, it is required that: 1) g(t) → r̂ when-
ever r(t) → r, with r̂ the best feasible approxima-
tion of r and θ(t) → 0m ; and 2) the ROG have a
finite settling time, viz. g(t) = r̂ and θ(t) = 0m for
a possibly large but finite t whenever the reference
stays constant after a finite time.
By linearity, one is allowed to separate the effects
of the initial conditions and inputs from those of
disturbances, e.g. x(t) = x(t) + x̃(t), where x(t) is
the disturbance-free component and x̃(t) depends
only on disturbances. Then, adopting the follow-
ing notations for the disturbance-free solutions of
(1) to a constant command z(t) = w, we define
xw := (In − Φ)−1 Gw
y w := Hy (In − Φ)−1 Gw (5)
cw := Hc (In − Φ)−1 Gw + Lw
Consider next the following set recursion
C0 := C ∼ Ld D
Ck := Ck−1 ∼ Hc Φk−1 Gd D
..
. (6)
∞

C∞ := Ck
k=0
where A ∼ E is defined as {a ∈ A : a +
e ∈ A, ∀e ∈ E}. It can be shown that the sets
Ck are nonconservative restrictions of C such that
c(t) ∈ C∞ , ∀t ∈ Z+ , implies that c(t) ∈ C, ∀t ∈
Z+ . Thus, one can consider only disturbance-free
134
evolutions of the system and adopt a ”worst-case”
approach. For reasons which have been clarified
in (Bemporad et al., 1997), it is convenient to
introduce the following sets for a given δ > 0,
C δ := C∞ ∼ Bδ and Wδ := w ∈ R2m : cw ∈ C δ ,
where Bδ is a ball of radius δ centered at the
origin. Let us assume that there exists a possibly
vanishing δ > 0 such that Wδ is non-empty. In
particular, Wδ is the set of all commands whose
corresponding steady-state solution satisfies the
constraints with margin δ. From the foregoing
is closed and convex.
The main idea is to choose at each time step a
constant virtual command z(·) ≡ w, with w ∈ Wδ
such that the corresponding virtual evolution ful-
fills the constraints over a semi-definite horizon;
trol law and the distance between the command
reference g(t) and the constant reference r(t) are
minimal. The ROG commands are applied, a new
state is measured and the procedure is repeated.
In this context we define the set V(x(t)) as
V(x(t)) = {w ∈ Wδ : c(k, x(t), w) ∈ Ck , ∀k ∈ Z+ }
(7)
where
k−1
c(k, x(t), w) = Hc Φk x(t) + Φk−i−1 Gw + Lw (8)
i=0
is to be understood as the disturbance-free virtual
evolution at time k of the constraints vector from
the initial condition x(t) at time zero under the
constant command z(·) ≡ w. As a consequence
V(x(t)) ⊂ Wδ , and, if non-empty, it represents
the set of all constant virtual sequences in Wδ
whose evolutions starting from x(t) satisfies the
constraints also during transients. It can also
be shown that such a set is finitely determined,
viz. there exists a positive integer k0 such that
(7) is identically characterizable by restricting
k ∈ {0, · · · , k0 }, with k0 computable off-line as
described in (Gilbert and Tin Tan, 1991).
The ROG output is based on the minimization of
a cost function subject to the constraints imposed
by (7). The cost function has the following form
J(x(t), z(t), r) = g(t) − r 2Ψg + θ(t) 2Ψθ (9)
where Ψg = ΨTg > 0m ,Ψθ = ΨTθ > 0m and
x Ψ := xT Ψx. Thus, at each time t ∈ Z+ , the
ROG output is chosen according to the solution
of the following constrained optimization problem
z(t) := arg min J(x(t), z, r) (10)
z∈V(x(t))
The following properties hold true for the above
described ROG.
Theorem 1. Let assumptions (A1) be fulfilled.
Consider system (3) along with the ROG selection
rule (10), and let V(x(0)) be non-empty. Then:
1. The minimizer in (10) uniquely exists at each The terminals ∆fref and θ respectively represent
t ∈ ZZ+ and can be obtained by solving a convex frequency reference and control offset for the two-
constrained optimization problem, viz. V(x(0)) area power system. All other variables are stan-
non-empty implies V(x(t)) non-empty along the dard and can be found in (Yang et al., 2002).
trajectories generated by the ROG command (10). Traditionally, LFC control consists in keeping the
2. The set V(x(t)), ∀x(t) ∈ IRn , is finitely system frequency deviation ∆f1 (t) in Area 1,
determined, viz. there exists an integer k0 such ∆f2 (t) in Area 2 and the deviation in the tie-line
that if c̄(k, x(t), w) ∈ Ck , k ∈ {0, 1, . . . k0 }, then power flow ∆Ptie (t) between the two areas as close
c̄(k, x(t), w) ∈ Ck ∀k ∈ ZZ+ . Such a constraint as possible to their nominal values despite of dis-
horizon k0 can be determined off-line. turbances ∆PD1 (t) and ∆PD2 (t). Accordingly the
3. The constraints are fulfilled for all t ∈ ZZ+ . nominal values of frequency references and control
4. The overall system is asymptotically stable; offsets are constant to zero. LFC requirements are
in particular, whenever r(t) ≡ r, lim θ(t) = 0m , usually satisfied by the so-called Automatic Load-
t→∞
and g(t) converges either to r or to its best steady- Frequency Control (ALFC) stage included in the
state admissible approximation r̂, with model (12), which is a suitable local controller

designed in a decentralized way, viz. by using
r̂ local information for feedback. See e.g. (Yang et
ẑ := := arg min J(x(t), z, r) (11)
0m z∈V(x(t)) al., 2002)-(Lim et al., 1997) and references therein
Consequently, by the offset-free condition (A1.2), for some decentralized design methods.
limt→+∞ ȳ(t) = r̂, where ȳ is the disturbance-free Here, besides standard LFC requirements, we
component of y. want also to satisfy the following set of constraints
Proof - The proof can be obtained following sim-
ilar ideas used in (Bemporad et al., 1997) and |∆fi (t)| ≤ ∆fmax,i , i = 1, 2, (13)
(Kolmanovsky and Sun , 2006). 2
|∆Ptie | ≤ βmax , (14)
|∆PTi (t)| ≤ γmax,i , i = 1, 2. (15)
3. POWER SYSTEM MODEL AND
at each time instant t. Specifically, constraints
PROBLEM FORMULATION
(13) prescribe bounds on the maximum frequency
deviations for each area, (14) limit the maximum
We shall consider a two-area power system for the
tie-line power flow, while (15) impose bounds on
LFC as depicted in Fig. 3 (Yang et al., 2002). A
the maximum generated power in the generators.
state space model is given by
In the sequel, we shall consider that the ALFC
control stage acts as the primal controller in the
_ +
ROG framework. Let
∆f1ref 
Bs1
1
R1 ∆PD1
 x(t + 1) = Āx(t) + B̄1 ∆fref (t) + B̄2 θ(t)+
+F̄ ∆PD (t)
∆Pc1
_ ∆Pv1 ∆PT1 _ 
+ -K1 1 1 Kp1 y(t) = C̄x(t)
s + + 1+sTG1 1+sTT1 + 1+sTp1
_ ∆f1
+
(16)
θ1
+ T12 ∆Ptie be the discrete-time form of equations (12) achieved
s
_ by sampling the system with a period Tc = 0.1 s.
a12 θ2 a12 Then, the system (16) together with the con-
∆Pc2 + ∆Pv2 ∆PT2 _
+ -K2 1 1 Kp2 ∆f2 straints (13)-(15) can be described as
+
s + _ 1+sTG2 1+sTT2
+ _ 1+sTp2

Bs2 1  x(t + 1) = Φx(t) + Gz(t) + Gd ∆PD (t)
R2 ∆PD2
y(t) = Hy x(t) (17)
∆f2ref

_ c(t) = Hc x(t) + Lz(t) + Ld ∆PD (t)
+

T
where z = [∆fref θT ]T , Φ = Ā, G = [B̄1 B̄2 ],
Gd = F̄ and Hy = C̄. Notice that the constrained
Fig. 3. Block diagram of a two-area power system vector c(t) is characterized by the following ma-
 trices
 ẋ(t) = Ax(t) + B1 ∆fref (t) + B2 θ(t)+  
C̄

+F ∆PD (t) (12) e5 Φ
y(t) = Cx(t) Hc =  
e2 Φ , Ld = 05×2 , L = 05×4 ,
where ∆fref = [∆f1ref ∆f2ref ]T , θ = [θ1 θ2 ]T , e7 Φ
y = [y1 y2 ]T = [∆f1 ∆f2 ]T , ∆PD = [∆PD1 ∆PD2 ]T ,
where ei is the canonical base of R9 . In particular,
x = [∆f1 ∆PT1 ∆Pv1 ∆Pc1 ∆Ptie ∆f2 ∆PT2 the vector c(t) ∈ R5 describes the dynamics of the
∆Pv2 ∆Pc2 ]T . constrained sequences (13)-(15).

135
 4. SIMULATIONS
The ROG constrained strategy of Section 2 has
been applied to the system (16). In the simula-
tions, we assumed δ = 10−6 , Ψg = I2×2 , Ψθ =
I2×2 . The initial conditions x(0) = 09 , z(0) = 04
were chosen, corresponding to the desired steady-
state. Note that the command vector z(t) includes
both the frequency and control offset set-points
for each area. The following constraints were con-
sidered
|∆f1 (t)| ≤ 3.167, |∆f2 (t)| ≤ 3.167,
|∆Ptie | ≤ 0.985, (18)
|∆PT1 (t)| ≤ 0.9, |∆PT2 (t)| ≤ 0.9.
A constraint horizon k0 = 300 was computed via
the numerical procedure given in (Gilbert and Tin
balance of the overall system cannot be ensured
under the prescribed constraints (see Figs. 7-8).
Nonetheless, the ROG unit is capable to mod-
ify the nominal frequency and control offset set-
points (see Fig. 6) by ensuring a new power bal-
ance. In fact, in response to this situation, the
Area 1 generator increases its power production,
a fraction of it is transferred to Area 2 via the
tie-line (5), so as to compensate the fault in Area
2 (Fig. 4, upper subplot on the right side). It is
worth commenting that the ROG is not informed
of the fault occurrence and its behaviour hinges
on its intrinsic reconfiguration capability. In such
a way the ROG is able to manage unpredictable
events.
∆ P [MW]
4
∆ f [Hz]
1
T

Tan, 1991). 2 0.5

The convex and compact set D, characterizing

AREA 1
0 0
load disturbances, has been defined as follows
  −2 −0.5

D := d ∈ R2 : U d ≤ h , (19)
    −4
0 50 100 150
−1
0 50 100 150

1 0 2
 0 1 2 4 1

where U =    
 −1 0  and h =  2  M W. AREA 2 2 0.5

0 −1 2 0 0

The above set includes all admissible load varia- −2 −0.5

tions. −4 −1
0 50 100 150 0 50 100 150
In order to test the effectiveness of the scheme, we Time (sec.) Time (sec.)

have considered two distinct critical events:

Fig. 4. Frequency (left side) and generating power
(a) a step load disturbance of ∆PD1 (t) = deviations constraints (right side): with ROG
1.65 M W acting to Area 1 from t = 5 s to
∆P [MW]
tie
t = 15 s; 1

(b) the Area 2 generator exhibits a power supply 0.8

of 30% loss (w.r.t. its nominal value) due 0.6

to a failure from t = 65 s to t = 100 s, 0.4

which implies that it is not able to satisfy 0.2

the nominal load demand. 0

In Figs. 4-5, the response of the plant under −0.2

the ROG action is reported and no constraints −0.4

violation occur. The simulations show that the −0.6

constraints become active when one of the two −0.8

events occurs. After each event, the system recov- −1

0 20 40 60 80 100 120 140 160
ers the steady-state condition with a settling time Time (sec.)

of nearly 50 s. In the next Fig. 6, it is worth noting

how the ROG modifies the frequency deviation
and control offset set-points from their zero nom-
inal values, allowing feasible system evolutions in
response to any disturbance within the set (19).
In the first part of the experiment a significant
load variation is simulated. When the ROG is
used (Figs. 4-5), if compared with the system
evolutions obtained by the ALFC stage acting
without ROG (Figs. 7-8), the resonance effect is
significantly damped. This can help to prevent
undesirable phenomena (e.g. generators out of ser-
vice, blackouts, etc.) when such anomalies occur.
The fault event (b) is more interesting, starting
from t = 65 s the Area 2 generator is not capable
to satisfy the load demand. Therefore the power
Fig. 5. Deviation in the tie-line power flow: with
ROG
Figs. 7-8 depict the response of the compensated
power system from the initial state x(0) = 09
without the ROG unit. Under the same scenario,
violations of the prescribed frequency and power
deviation constraints come out. It turns out that
the single action of the local controllers is not ca-
pable to enforce the constraints during transients.
5. CONCLUSIONS
A discrete-time predictive strategy for constrained
load frequency control problem in a two-area in-

136
 ROG frequency commands ROG offset commands events. In particular when a loss of power pro-
1.5 3

1
2.5 duction occurs, the proposed strategy is capable
0.5
2
to modify the frequency set-point and the offset
1.5

Area1 0 1 on the control law in such a way that constraints

-0.5
0.5
fulfilment is achieved.
0

-1 -0.5
0 50 100 150 0 50 100 150

1.5 0.5 REFERENCES

0
1
-0.5
Andersson G. (2004). Modelling and analysis of
-1
Area2

0.5
-1.5 eletrical power systems. EEH-Power Systems
0 -2
Laboratory, ETH Zürich.
-2.5
-0.5 3
Bemporad A., Casavola A. and Mosca E.
0 50 100 150 0 50 100 150

Time (sec.) Time (sec.)

(1997). “Nonlinear Control of Constrained
Linear Systems via Predictive Reference Man-
Fig. 6. Computed commands z(t) : frequency agement”. IEEE Trans. Automat. Control,
deviations [Hz] (left side) and control offsets 42, 340–349.
(right side) Bhowmik S., Tomsovic K. and Bose A. (2004).
“Communication models for third party load
∆ P [MW]
6
∆ f [Hz]
2
T frequency control”. IEEE Transactions on
4 1.5 Power Systems, 19(1), 543–548.
1
Casavola A., Papini M. and Franzè G. (2006).
AREA 1

2
0.5
0
0
“Constrained supervision of dynamic systems
−2
−0.5 in spatial networks”. IIEEE Transaction on
−4 −1
0 50 100 150 0 50 100 150 Automatic Control, 51(3), 421–437.
6 1
Casavola A., Mosca E. and Angeli D. (2000). “Ro-
4 0.5
bust command governors for constrained lin-
ear systems”. IEEE Trans. Automat. Control,
AREA 2

2 0

0 −0.5 45, 2071–2077.

−2 −1 Çimen H. (1998). “Decentralised load frequency
−4
0 50 100 150
−1.5
0 50 100 150 controller design based on structured singular
Time (sec.) Time (sec.)
values”. IEE Proc.-Gener. Transm. Distrib.,
Fig. 7. Frequency (left side) and generating power 145(1), 7–14.
deviations constraints (right side): without Gilbert E.G. and Tin Tan K. (1991). “Linear sys-
ROG tems with state and control constraints: The
theory and applications of maximal output
∆ Ptie [MW]
1.5
admissible sets”. IEEE Trans. Automat. Con-
trol, 36, 1008–1020.
1
Kolmanovsky I. and Sun J. (2006). “Parameter
Governors for discrete-time nonlinear systems
0.5
with pointwise-in-time state and control con-
straints”. Automatica, Vol. 42(5), 2006, pp.
841-848.
0
Lim K.Y., Wang Y. and Zhou R. (1997). “Decen-
tralised robust load-frequency control in coor-
−0.5
dination with frequency-controllable HVDC
links”. Electrical Power and Energy Systems,
−1
0 20 40 60 80
Time (sec.)
100 120 140 160 19(7), 423–431.
Nobile E., Bose A. and Tomsovic K. (2001). “Fea-
Fig. 8. Deviation in the tie-line power flow: with- sibility of a bilateral market for load follow-
out ROG ing”. IEEE Transactions on Power Systems,
16, 782–787.
terconnected power system has been proposed. Yang T.C., Ding Z.T. and Yu H. (2002). “De-
Constraints on frequency and power deviations centralised power system load frequency con-
during transients have been enforced. All the trol beyond the limit of diagonal domi-
relevant concepts of the control strategy have nance”. Electrical Power and Energy Systems,
been discussed and summarized. Simulation re- 24, 173–184.
sults have shown that the proposed approach en- Wang Y., Zhou R. and Wen C. (1993). “Robust
sures viable evolutions to the overall power sys- load-frequancy controller for power systems”.
tem with respect to the prescribed operative con- IEE Proceedings-C, 140(1), 11–16.
straints despite changes in the loads and failure

137