A Benchmark Problem for Response

Control of Wind-Excited Tall

Buildings
Jann N. Yang1, Jong-Cheng Wu2, Bijan Samali3, and Anil K. Agrawal1
1 Department of Civil & Environmental Engineering, University of California, Irvine, CA, U.S.A.
2 Department of Civil Engineering, Tamkang University, Taiwan, R.O.C.
3 Department of Civil Engineering, The University of Technology, Sydney, Australia

ABSTRACT

This paper presents an overview and problem definition of a benchmark problem for the
response control of wind-excited tall buildings. The building considered is a 76-story 306
meters concrete office tower proposed for the city of Melbourne, Australia. The building is
slender with a height to width ratio of 7.3; hence, it is wind sensitive. Either active, semi-
active or passive control systems can be installed in the building to reduce the wind
response, although only an active control sample problem has been worked out to illustrate
the control design. In the case of active control systems, either an active tuned mass damper
(ATMD) or an active mass driver (AMD) can be installed on the top floor. In the case of
passive or semi-active systems, such as viscous dampers, visco-elastic dampers, ER/MR
dampers, etc., control devices can be installed in selected story units. Control constraints
and evaluation criteria are presented for the design problem. A simulation program based on
the LQG has been developed and made available for the comparison of the performance of
various control strategies.

1. INTRODUCTION

Significant progress has been made in active structural control against natural hazards, such
as earthquakes and strong winds. In particular, active control systems have been
implemented in full-scale buildings in Japan to alleviate the acceleration response under
wind excitations [e.g., Housner and Masri (1994)]. While various control algorithms have
been explored for applications to active response control of tall buildings and towers subject
to wind loads [e.g., Yang and Samali (1983), Samali et al (1985), Kareem (1981), Suhardjo
et al (1992), Ankireddi and Yang (1996), Wu and Yang (1997a, b), Cao et al (1997), Wu et
al (1998), Yang et al (1997a)], no systematic study has been conducted to evaluate the
performances of different control strategies based on the same criteria using the same
structural model under the same wind loads. To focus the future efforts in the most effective
direction, two pilot benchmark problems had been proposed for active response control
against earthquakes using laboratory experimental structural models [Spencer et al (1998)].
Based on realistic full-scale buildings, a series of two active structural control benchmark
problems have been developed for earthquake and wind excitations, respectively, and this
paper is the second one of such a series.
Wind-excited tall buildings involve excessive degrees of freedom so that the
computational effort involved in determining the dynamic response becomes excessive. As
a result, a reduced-order model has been constructed, referred to as the evaluation model,
from which the structural response quantities are computed and evaluated. To reduce the
computational efforts, the selected benchmark building is symmetric in both horizontal
directions, and the axis of elastic centers coincides with the axis of mass centers to avoid the
coupled lateral-torsional motions [e.g., Yang and Samali (1983), Samali et al (1985)]. Under
wind excitations, the building undergoes both the along-wind and across-wind motions, so
that multiple active mass dampers (controllers) [e.g., Housner and Masri (1994)] or a single
mass damper with multiple actuators in two directions [e.g., Samali et al (1985), Cao et al
(1997)] are needed. To simplify the computational efforts, we only consider the along-wind
motion so that only one active mass damper with a single actuator is needed, although in
reality, the combined response quantities due to both along-wind and across-wind are larger.
Further, the actuator dynamics and controller-structure interaction have been neglected in the
sample controller design to be given later, although these effects have been observed to be
significant [Dyke et al (1995)]. The main justification for making so many simplifications
above is that our main purpose of the benchmark study is to evaluate and compare the
performance of various control strategies. More sophisticated benchmark problems, which
will involve more computational efforts, should be considered after more experience is
gained through the present study.

2. 76-STORY BUILDING AND MODEL

The building considered is a 76-story 306 meters office tower proposed for the city of
Melbourne, Australia as shown in Figs. 1 and 2. This is a reinforced concrete building
consisting of concrete core and concrete frame. The core was designed to resist the majority
of wind loads whereas the frame was designed to primarily carry the gravitational loads and
part of the wind loads. All the relevant structural analyses and design had been completed;
however, it was not built due to economic recession. The building has a square cross-section
with chamfer at two corners as shown in Fig. 1. The total mass of the building, including
heavy machinery in the plant rooms, is 153,000 tons. The total volume of the building is
510,000 m3, resulting in a mass density of 300 kg per cubic meter, which is typical of
concrete structures. The building is slender with a height to width ratio (aspect ratio) of
306.1/42= 7.3; therefore, it is wind sensitive.
The perimeter dimension for the center RC core is 21m x 21m. The reinforced concrete
perimeter frame consists of columns at 6.5 m, which are connected to a 900mm deep and
400mm wide spandrel beam on each floor. There are 24 perimeter columns on each level
with 6 columns on each side of the building. The lightweight floor construction uses steel
beams with metal deck and a 120mm slab. The compressive strength of concrete is 60 MPa
and the modulus of elasticity is 40 GPa. Column sizes, core wall thickness and floor mass
vary along the height, and the building has six plant rooms.
The 76-story tall building is modelled as a vertical cantilever beam. A finite element
model is constructed by considering the portion of the building between two adjacent floors
as a classical beam element of uniform thickness, leading to 76 translational and 76
rotational degrees of freedom. Then, all the 76 rotational degrees of freedom have been
removed by the static condensation. This results in a 76 degrees of freedom, representing
the displacement of each floor in the lateral direction. The first five natural frequencies are
0.16, 0.765, 1.992, 3.790 and 6.395 Hz. The first three mode shapes are shown in Fig. 3.
Damping ratios for the first five modes are assumed to be 1% for the proportional damping
matrix. This model, having (76 ×76) mass, damping and stiffness matrices, is referred to as
the “76 DOF Model”.
A tuned mass damper (TMD) with an inertial mass of 765 tons is installed on the top
floor. This is about 68.9% of the top floor mass, which is 0.5% of the total mass of the
building. The natural frequency is tuned to 0.16 Hz and a damping ratio of 20%, which is
higher than the optimal one, is used. A higher damping is used in order to allow for a bigger
stroke for the active tuned mass damper (ATMD). The building with ATMD is referred to
as the “77 DOF Model”. As will be described later, the designer can choose appropriate
natural frequency and damping ratio for the ATMD system.

Figure 1. Plan View of the 76-Story Building.

 Figure 2. Elevation View of the Building.

2.1 Simplified Evaluation Model

The numerical computation of the response quantities, including root-mean square (rms)
response, power spectral density, etc., is very time-consuming and computationally
expensive for the 76 DOF and 77 DOF models. For example, the computation of rms
responses for the 76 DOF model using MATLAB takes about 8-9 hours on the SUN Sparc
20 workstation. Hence, a state order reduction method has been used to derive lower order
models for the building with and without ATMD. In this method, a state reduced-order
system is derived such that the eigen properties of the selected modes as well as the selected
states of the original system are represented in the reduced-order system [e.g., Davison
(1966), Wu et al (1998)]. Based on this approach, the 77 DOF model (with ATMD) is
reduced to a 24 DOF system such that the first 48 complex modes (eigenvalues and
eigenvectors) of the 77 DOF system are retained. The resulting state equation is given by
x& = Ax + Bu + EW (1)
, x&′
in which x = [x ′ ]′is the 48-dimensional state vector, u is the scalar control force, and W
 76 76 76

0
Mode 1 ( f= 0.160 Hz ) Mode 2 ( f = 0.765 Hz ) Mode 3 ( f = 1.992 Hz )
Figure 3. Modeshapes of First Three Modes of the Building.

is the wind excitation vector of dimension 77. In Eq. (1) above, x = [ x 3 , x 6 , x10 , x13 ,
x16 , x 20 , x 23 , x 26 , x 30 , x 33 , x 36 , x 40 , x 43 , x 46 , x 50 , x 53 , x 56 , x 60 , x 63 ,
x 66 , x 70 , x 73 , x 76 , x m ]′
, where x i is the absolute displacement of the i-th floor and
x m is the relative displacement of the inertial mass of damper with respect to the top floor.
Also, A is a ( 48 ×48) system matrix, B is a 48 location vector, and E is a ( 48 ×77 ) matrix.
Similarly, the 76 DOF model (building without ATMD) can be reduced to a 23 DOF
system by retaining the first 46 complex modes of the original system. The resulting state
equation is also expressed by Eq. (1). In this case, however, the dimensions of x, A, B, and
E are ( 46 ×1) , ( 46 ×46) , ( 46 ×1) and ( 46 ×76) , respectively, and u=0. Matrices A, B, and
E are completely different from those for the 24 DOF system. Despite the system reduction
above, the computer time required for the response quantities is still too long, because the
wind loads acting on each of the 23 floors above are computed from the wind loads W acting
on each of the 76 floors through a transformation. For the stochastic response analysis (see
Appendix I), the cross-power spectral density matrix of the wind loads W is (76 ×76) . Each
element of such a big matrix has to be evaluated at every frequency ω and then multiplied by
a transformation matrix as well as the frequency response (transfer function) matrix in order
to obtain the response power spectral density matrix of the controlled output vector. The
required number of operations is excessive even for a high speed computer.
To further reduce the computational efforts, instead of reducing wind loads through the
model reduction method (transformation) described above, the wind load vector W is
modelled physically by lumping wind loads on adjacent floors at the locations that
correspond to the 24 DOF model (or 23 DOF model). Thus, in Eq.(1), the dimension of W
becomes 24 and E becomes a ( 48 ×24) matrix. These simplified models are denoted by 24
DOF with W24 and 23 DOF with W23 models, respectively. As a result of this
simplification, the computational time for the response quantities is reduced to be
reasonable.
 To verify the accuracy of different reduced-order models above for the building,
numerical simulations were conducted to calculate the rms and the power spectral density
responses of various floors. The rms response quantities for various models subject to the
wind loads used for the sample controller to be described later are shown in Table 1. In
Table 1, σx i and σ&x&i represent the rms values of the floor absolute displacement x i and
absolute acceleration &x&i , respectively, and σxm and σ&x&m are the rms values of the relative
displacement x m and absolute acceleration & x&m of the mass damper, respectively. It is
observed from Table 1 that the rms response quantities for the 23 DOF and 24 DOF models
are very close to those for the 76 DOF and 77 DOF models, respectively. Consequently, the
reduced-order structure is an adequate representation of the full-order structure (76 DOF).
An mentioned previously, the dimension of wind loads W used for the response computation
is 76 in this case.

Table 1. Stochastic Response Quantities of the 76-Story Building Using Different

Simplified Models.

76 DOF Model 77 DOF Model

Floor σxi σ&x&i σx i σ&x&i
No. 2
cm cm/s cm cm/s2
(1)
(2) (3) (4) (5)
1 0.0161 0.1049 0.0094 0.1036
30 2.0830 2.2705 1.2108 1.4441
50 5.0632 4.8033 2.9414 2.3377
75 9.6611 9.2286 5.6134 4.5741
76 9.8791 9.5122 5.7402 4.8220
md - - 9.4569 10.2732
23 DOF Model 24 DOF Model
Floor σxi σ&x&i σx i σ&x&i
No.
cm cm/s2 cm cm/s2
(6)
(7) (8) (9) (10)
1 0.0161 0.2413 0.0094 0.2408
30 2.0830 2.4176 1.2108 1.6660
50 5.0632 4.9668 2.9414 2.6577
75 9.6611 9.4023 5.6134 4.9160
76 9.8791 9.6669 5.7402 5.1196
md - - 9.4569 10.2732
23 DOF Model with W23 24 DOF Model with W24
Floor σxi σ&x&i σx i σ&x&i
No.
cm cm/s2 cm cm/s2
(11)
(12) (13) (14) (15)
1 0.0163 0.2558 0.0094 0.2549
30 2.1167 2.7153 1.2159 2.0136
50 5.1447 5.1109 2.9531 2.7658
75 9.8184 9.9137 5.6389 5.5464
76 10.0400 10.3438 5.7664 6.0202
md - - 9.6555 10.4917
 It is further observed from Table 1 that the displacement response quantities for 23 DOF
with W23 and 24 DOF with W24 models are quite close to those for 76 DOF and 77 DOF
models. However, there are discrepancies for the absolute acceleration response. Such a
discrepancy comes from the physical modelling (lumping) of wind loads rather than the
model reduction of structures. Nevertheless, the 23 DOF with W23 and 24 DOF with W24
models will be considered to be reasonable for the evaluation model in Eq. (1), since our
purpose is to investigate the performance of various control strategies.
The controlled output vector z and the measured output vector y of the evaluation model
in Eq. (1) can be expressed as
z = C z x + D z u + Fz W (2)
y = C y x + D y u + Fy W + v (3)
in which z = [ x1 , x 30 , x 50 , x 55 , x 60 , x 65 , x 70 , x 75 , x 76 , x m , x&1 , x&30 , x&50 , x&55 ,
x&60 , x&65 , x&70 , x&75 , x&76 , x&m , & x&1 , &
x&30 , &
x&50 , &
x&55 , &x&60 , &
x&65 , &x&70 , &
x&75 , &
x&76 ,
&&m ]′is a 30-dimensional vector of the response that can be regulated, and y =[ x&1 ,
x x&30 ,
x&50 , x&55 , x&60 , x&65 , x&70 , x&75 , x&76 , x&m , &
x&1 , &
x&30 , &
x&50 , &
x&55 , &
x&60 , &
x&65 , &
x&70 , x
&&75 ,
x&76 , &x&m ]′is a 20-dimensional vector of the response that can be directly measured by
&
installing sensors. However, the maximum number of sensors that can be used will be
specified later. In Eq. (3), v is a vector of measured noises, and C z , D z , Fz , C y , D y ,
and Fy are matrices of appropriate dimensions. In the notation above, x&m is the relative
velocity of the mass damper with respect to the top floor.
The model given in Eqs. (1)-(3) for 24 DOF with W24 is termed as the evaluation model
and it will be used to assess the performance of candidate control strategies, i.e., the
evaluation model is considered to be the true representation of the 76-story building
equipped with an ATMD.

3. CONTROLLER DESIGN PROBLEM

The problem of controller design is to determine a discrete-time feedback compensator of

the form
x c (k + 1) = f1[x c (k), ~
y (k), u(k), k] (4)
~
u(k) = f 2[x c (k),y(k),k] (5)
~
where x c (k) , y(k) and u ( k ) are vectors for the states of the compensator, selected
measurement output vector, and the control force, respectively, at time t= k∆T with ∆T
being the sampling time for the compensator. The maximum dimension of the compensator
is required to be less than 12, i.e., dim [ x c (k) ] ≤12. For each proposed control design, the
performance should be evaluated using the evaluation model 24 DOF with 24W.

4. WIND EXCITATION MODEL

The wind velocity can be decomposed into an average wind velocity and a wind
fluctuation component, that essentially is a stationary random process. The average wind
velocity varies along the height of the building. Although the wind load depends on the
response of the structure, such an interaction effect can usually be neglected and the wind
load can be computed from the wind velocity. Consequently, the wind load consists of a
static load due to the average wind velocity and a dynamic load due to wind velocity
fluctuations. For structural control, only the fluctuating wind loads will be considered.
Since the building is symmetric in both horizontal directions and since the axis of elastic
centers and the axis of mass centers coincide, there is no coupled lateral-torsional motions.
Further, to reduce the computational efforts, only the along-wind motion will be considered
in this study, although the combined response quantities due to along-wind and across-wind
excitations are larger. This simplification is made because the main purpose of the
benchmark study is to evaluate and compare the performance of different control strategies.
The well-known Davenport wind load spectrum, which has been used in the Canadian
design code is used herein for along-wind loads for simplicity. The (i, j) element of the two-
sided cross-power spectral density matrix S ww (ω ) of the wind load in the along-wind
motion can be expressed as [Simiu and Scanlan (1986)]
8 w i w j K 0 V 2r (600ω π V r )2  c1 | ω | | h i - h j | 
S w w (ω ) = exp 
[ ] - 2π (6)
i j
V i V j | ω | 1 + (600ω π )2 4/3
 Vr  
Vr
where ω is in radian per second,
2
w j = E[w j (t)] = 0.5 ρ A j C D V j (7)
is the average wind force (static force) on the jth floor, Vr is the reference mean wind
velocity in meters per second at 10 m above the ground, c1 is a constant, h i is the height of
the i-th floor, and K0 is a constant depending on the surface roughness of the ground [Simiu
and Scanlan (1986)]. In Eq. (7), ρ is the air density, C D is the drag coefficient, A j is the
tributary area for the j-th story unit. The mean wind velocity V i is assumed to follow a
power law
α
Vi = Vg ( h i h g ) (8)
in which h g is the gradient height, Vg is the average wind velocity at the gradient height,
and α is a constant between 0.15 and 0.5. The values for the parameters h g and α can be
found in [Simiu and Scanlan (1986)], which are characterized by the ground condition, such
as the roughness.
A sample function of the dynamic wind load on the jth floor can be simulated as follows
[Shinozuka (1985)]
j N
w j (t) = ∑ ∑ 2 ∆ω | H jm ( ω n ) | cos[ω n t + θ jm ( ω n ) + φmn] (9)
m =1 n =1
where Hjm is the ( j, m) element of a lower triangular matrix H(ω ) given by
S ww ( ω ) = H( ω ) H (ω ) ′ (10)
and H (ω ) is the complex conjugate of H(ω ) , N is the total number of sampling frequencies
and ω n = (n-1) ∆ω with ∆ω being the sampling frequency. The phase angle θ jm (ω n ) is
defined as
 Im H jm ( ω n ) 
θ jm ( ω n ) = tan -1   (11)
Re H jm ( ω n ) 
 
and φmn are statistically independent random variables uniformly distributed between 0 and
2π.
For the benchmark problem, the following parameters for the cross-power spectral
density, Eqs.(6)-(8), are used: K0 = 0.03, c1 = 7.7, α = 0.4, air density ρ = 1.25 kg/m3, drag
coefficient CD = 1.2, and gradient height h g = 300 m. For the evaluation model in Eq.(1),
the tributary areas for each of the 24 DOF W24 are 961.8, 655.2, 491.4, 491.4, 655.2, 491.4,
655.2, 491.4, 491.4, 491.4, 655.2, 541.8, 491.4, 655.2, 491.4, 491.4, 655.2, 491.4, 491.4,
655.2, 491.4, 491.4, 378.0, and 0 m2, respectively, whereas the heights for each wind load
location are 19.0, 30.7, 46.3, 58.0, 69.7, 85.3, 97.0, 108.7, 124.3, 136.0, 147.7, 164.5, 176.2,
187.9, 203.5, 215.2, 226.9, 242.5, 254.2, 265.9, 281.5, 293.2, and 306.1 m, respectively.
The tributary area for ATMD has been assumed to be zero and the wind force acting on
ATMD is zero. The mean wind velocity at 10 m above the ground is chosen to be 15 m/sec.,
i.e., Vr = 15 m/sec., reflecting the annual maximum wind velocity in the region of moderate
to active wind environments.
Further, a set of sample functions w j ( t ) of wind loads acting on each DOF has been
simulated using Eqs. (9-11) with ∆ω =0.02 rad/sec., N=1024, and a duration of about 314
sec. However, only a duration of 300 sec. is used for the integration to determine the
structural response. This set of sample functions (time histories) of wind loads is also
available in Yang et al (1997b) for the deterministic analysis of the building response.

5. PERFORMANCE CRITERIA AND CONSTRAINTS

5.1 Performance Criteria: RMS Responses Based on Stochastic Analysis

Using the cross-power spectral density matrix in Eq.(6) and the random vibration analysis
described in Appendix I, the rms displacement and acceleration responses can be evaluated
for the proposed design of controller. Since the calculation of stochastic response quantities
for all the floors may require too much computational time, only a calculation of the rms
displacement and acceleration for 9 floors contained in the controlled output vector z is
required. Since the main objective of installing an active control system on top of the 76-
story building is to reduce the absolute acceleration to alleviate the occupant’s discomfort,
the first evaluation criterion for the controllers is their ability to reduce the maximum floor
acceleration. A nondimensionalized version of this performance criterion is given by
 σ σ σ σ σ σ σ σ 
J 1 = max  &x&1 , &x&30 , &x&50 , &x&55 , &x&60 , &x&65 , &x&70 , &x&75  (12)
 σ&x&75o σ&x&75o σ&x&75o σ&x&75o σ&x&75o σ&x&75o σ&x&75o σ&x&75o 
where σ&x&i is the rms acceleration of the ith floor, and σ&x&75o = 9.914 cm / sec 2 is the rms
acceleration of the 75th floor without control, see Table 3. In the performance criterion J1 ,
accelerations only up to 75th floor are considered because the 76th floor is the top of the
building and it is not used by the occupants.
 The second criterion is the average percentage of acceleration reduction for floors above
the 49th floor, i.e.,
1
J 2 = ∑ [( σ&x&io − σ&x&i ) σ&x&io ] ; for i= 50, 55, 60, 65,70 and 75 (13)
6 i
in which σ&x&io is the rms acceleration of the ith floor without control. The third and fourth
evaluation criteria are the ability of the controllers to reduce the floor displacements. The
normalized version are given as follows
 σ σ σ σ σ σ σ σ σ 
J 3 = max x1 , x 30 , x50 , x55 , x60 , x65 , x70 , x75 , x76  (14)
 σx76o σx 76o σx76o σx 76o σx 76o σx 76o σx 76o σx 76o σx 76o 
1
J4 = ∑ [( σxio − σxi ) σxio ] ; for i= 50, 55, 60, 65,70, 75 and 76 (15)
7 i
where σxi and σxio are the rms displacements of the ith floor with and without control,
respectively, and σx 76o = 10.040 cm is the rms displacement of the 76 floor of the
uncontrolled building, see Table 3.
Each proposed control design must satisfy the actuator capacity constraints given by
σu ≤100 kN and σxm ≤25 cm , where σu and σxm are rms control force and rms
actuator stroke, respectively. In addition to above constraints, the control effort
requirements of a proposed control design should be evaluated in terms of the following
nondimensionalized criteria
J 5 = σxm σx 76o (16)
J 6 = σx&m σx&76o (17)
in which σx&m is the rms actuator velocity (relative velocity between the ATMD and the top
floor). The performance criteria in Eqs.(16) and (17) correspond to the physical size (i.e.,
stroke) and control power (i.e., actuator velocity) of the actuator. For the building without
control, σx&76o is 9.328 cm/sec.

5.2 Performance Criteria: Peak Responses Based on Deterministic Analysis

A set of 23 sample functions (time histories) of wind loads acting on each floor of the
evaluation model has been simulated and is available in Yang et al (1997b). A numerical
simulation (integration) for the on-line implementation of the proposed control design should
be conducted to evaluate the performance in terms of the following nondimensionalized
criteria

 & x&p1 &
x&p30 & x&p50 & x&p55 & x&p60 &x&p 65 & x&p70 &x&p75 

J 7 = max  , , , , , , ,  (18)
 &
& &
& &
& &
& &
& &
& &
& &
&
 p75o p 75o p75o p75o p75o p 75o p75o p75o 
x x x x x x x x 
1
J 8 = ∑ [( &x&pio − &x&pi ) &
x&pio ] ; for i= 50, 55, 60, 65,70 and 75 (19)
6 i

 x p1 x p30 x p50 x p55 x p60 x p 65 x p 70 x p75 x p76  
J 9 = max  , , , , , , , ,  (20)

 x p76o x p 76o x p76o x p76o x p 76o x p76o x p76o x p 76o x p76o 

 1
J10 = ∑ [(x pio − x pi ) x pio ] ; for i= 50, 55, 60, 65,70, 75 and 76 (21)
7 i
&&pi = peak acceleration of ith floor, x pio = peak
where x pi = peak displacement of ith floor, x
displacement of ith floor without control, and & x&pio = peak acceleration of ith floor without
control; for instance, x p76o = 26.009 cm and &x&p75o = 26.334 cm/sec2, see Table 4.
The actuator capacity constraints for the deterministic response analysis are: the
maximum control force max u( t ) ≤300 kN and the maximum stroke max x m ( t ) ≤ 75 cm.
In addition, the proposed control designs should be evaluated for the following control
capacity criterion
J11 = x pm x p76o (22)
J12 = x&pm x&p76o (23)
where x pm = peak stroke of actuator, x&pm = peak velocity of the actuator, and x&p76o =
22.622 cm/sec = peak velocity of 76th floor without control.

5.3 Design Constraints For ATMD

In order to make the benchmark problem as realistic as possible, the following

implementation constraints are imposed on the proposed control design:
1. In principle, acceleration and velocity of all floors can be available from measurement
for feedback to determine the control action. However, for convenience of benchmark
comparison, only the 20 variables in the measured output y defined by Eq. (3) can be
available. Further, a maximum of 6 sensors is imposed. In other words, one has the
degree of freedom to choose at most 6 variables in y defined by Eq. (3) as the feedback
quantities for the determination of the control action.
2. Although velocity sensors are available, the accuracy of measured velocities may not be
satisfactory. In this case, the velocity feedback may be obtained by passing the
measured acceleration through a filter as described in Spencer et al (1998).
3. The controller for the structure is digitally implemented with a sampling time of
∆t=0.001 sec.
4. A computational time-delay of 1000 µ sec. is assumed in the implementation of the
proposed controller. This time-delay is considered only in the deterministic response
analysis using simulated sample functions of wind loads.
5. The measurement noises are modelled as Gaussian rectangular pulse processes with a
pulse width of 0.002 seconds and a two-sided spectral density of 10 − 9 m2/sec3/Hz.
6. To limit the computational resources, the compensator for the controller in Eqs. (4) and
(5) is restricted to have no more than 12 states.
7. The performance of each control design should be evaluated using the evaluation model,
i.e., the 24 DOF with W24 model in Eq.(1).
8. The controller in Eqs.(4)-(5) is required to be stable.
9. The natural frequency and damping ratio of ATMD (or TMD) are design parameters
that can be chosen by the designer.
10. The robustness of the controller should be discussed.
 The Simulink model shown in Fig. 4 has been developed to simulate the features and
limitations of the structural control problem above and to compute both the stochastic and
deterministic response quantities. A time step of ∆t = 0.001 sec. for the integration has been
used in the simulation.

Figure 4. Simulink Model of the Building with LQG Controller.

6. SAMPLE CONTROLLER DESIGN

To illustrate some challenges of the benchmark problem, a controller design based on the
linear quadratic Gaussian (LQG) theory is presented. In the LQG approach, a reduced-order
system is constructed for the design of the controller, referred to as the control design model,
as follows
x&r = A r x& r + B r u + E r W (24)
z r = Czr x r + D zr u + Fzr W (25)
y r = C yr x r + D yr u + Fyr W + v r (26)
in which x r = [ x16 , x 30 , x 46 , x 60 , x 76 , x m , x&16 , x&30 , x&46 , x&60 , x&76 , x&m ]′= a
12-dimensional state vector, z r = a 30-dimensional controlled output vector identical of z
defined by Eq.(2), y r = [&x&50, & x&m ]′
x&76 , & = a 3-dimensional measured output vector, v r = a
3-dimensional measurement noise, and C zr , D zr , Fzr , C yr , D yr , and Fyr are
appropriate matrices. Here, the state order reduction technique [e.g., Davison (1966), Wu et
al (1998)] has been used by retaining the first 12 complex modes of the evaluation model. In
Eqs.(24)-(26), the excitation W and the measurement noise v r are assumed to be
uncorrelated Gaussian white noise vector processes. However, different components within
W can be correlated.
A state feedback LQG controller is obtained by minimizing the quadratic objective
function
J = lim E  ∫τ0 ( z ′ Q z + Ru 2 ) d t 
1
(27)
τ→ ∞ τ   

in which z = z r − Fzr W = C zr x r + D zr u , Q = a ( 30 ×30) diagonal weighting matrix =
diag [ 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 105 , 105 , 105 , 105 , 0, 1, 1, 1, 1, 1, 105 , 105 ,
105 , 105 , 0] and R = 3 ×10 − 2 . Minimizing the objective function in Eq.(27), the optimal
controller is obtained as [e.g., Skelton (1988)].
u = K x r ; K = − R − 1 ( B′
r Pc + S ′
) (28)
where Pc is the solution of the Riccati matrix equation
Pc A+A ′Pc − Pc B r R − 1 B ′
r Pc + Q − S R
− 1 S′
=0 (29)
and Q = C ′ − 1 S′
zr Q C zr , R = D ′
zr Q D zr + R , S = C ′
zr Q D zr and A = A r − B r R .
The controller obtained in Eq. (28) requires the reduced-order state feedback x r ,
which can be estimated from an observer, denoted by x̂ r , based on the separation principle.
The calculation of K can be made using the MATLAB function lqry within the control
toolbox. The Kalman-Bucy filter [e. g., Wu et al (1998), Skelton (1988)] can be designed to
estimate x̂ r as follows
x̂& = A x̂ + B u + L( y − C x̂ − D u)
r r r r r yr r yr (30)
in which the observer gain matrix L in Eq. (30) is obtained from
−1
L = ( Po C ′
yr + S o ) R o (31)
in which Po is the solution of the Riccati matrix equation
−1 −1
Po A o + A ′
o Po − Po C′
yr R o C yr Po + Q o − S o R o S′o =0 (32)
and
−1
A o = A′ r − C′yr R o S ′ o (33)
In Eqs.(31)-(33), Q o and R o are the auto-power spectral density matrices of two vectors
E r W and Fyr W + v r , respectively, and S o is the cross-power spectral density of two

vectors E r W and Fyr W + v r . These matrices are given by

Q o = E r SWW E ′ r ; So = E r SWW Fyr ′ ; R o = Sv v + Fyr SWW Fyr

r r
′ (34)
where SWW and Sv r v r are the power spectral density matrices of the white noises W and
v r , respectively, for the design purpose. For the observer design, SWW and Sv r v r can be
scaled appropriately for convenience of numerical computations. For this sample controller,
we choose Sv r v r = diag [0.1, 0.1, 0.1] and SWW =2000· S ww (ω ) at ω = 0.79 rad/sec,
where S ww (ω ) is the cross-power spectral density matrix of the wind load whose (i, j)
element is defined by Eq.(6). Calculation of L can be made using the MATLAB subroutine
lqew within the control tool box.
Finally the controller in Eqs .(28)-(30) is converted to the form in Eqs. (4)-(5) as follows
x c (k + 1) = A c x c (k) + B c y r (k) (35
u(k) = C c x c (k) + D c y r (k) (36)
in which dim[x c ( k )] = 12 and x c = x r = x̂ r .
Based on the eigenvalue analysis of the evaluation model, both the controller and the
closed-loop system are found to be stable. To assess the performance of the controller, both
stochastic and deterministic response analyses have been conducted using the computer
programs available at the website [Yang et al (1997b)]. For the stochastic response analysis,
the rms values and the power spectral densities of the response quantities are determined
directly from the cross-power spectral density matrix S ww (ω ) in Eq.(6). The method of
stochastic response analysis is presented in Appendix I. For the deterministic response
analysis, time histories of the response quantities are determined from the simulated set of
sample functions of wind loads acting on every floor. It should be mentioned that the
simulated sample functions of the wind loads are stationary processes but not ergodic, and
they are subject to significant sampling fluctuations. Hence, it is necessary to carry out the
stochastic response analysis to determine the rms response. Based on our experience, the
required computer time for the stochastic analysis is much less than that for the deterministic
response analysis.
The results for the evaluation criteria are presented in Columns (4) and (6) of Table 2, in
which the quantities in the parenthesis correspond to the building equipped with the TMD
(no actuator). In addition, the required active control force and the stroke of the actuator are
also shown in Table 2. As observed from Table 2, the hard constraints on the actuator
requirements, i.e., σu ≤100 kN , σxm ≤25 cm , max u( t ) ≤300 kN and max x m ( t ) ≤75
cm, are satisfied.

Table 2. Evaluation criteria for the sample controller.

Stochastic Response Analysis Deterministic Response

Analysis
Eval. ∆K = +15% ∆K = -15 % ∆K = 0% Eval. ∆K = 0%
Criteria (2) (3) (4) Criteria (6)
(1) (5)
J1 0.478 0.416 0.369 (0.559) J7 0.481 (0.803)
J2 0.483 0.548 0.606 (0.474 ) J8 0.460 (0.229)
J3 0.647 0.572 0.509 (0.574) J9 0.569 (0.652)
J4 0.353 0.429 0.491 (0.426 ) J10 0.423 (0.355 )
J5 2.277 2.202 1.870 (0.962) J11 2.401 (1.290)
J6 2.429 2.267 1.891 (1.030) J12 2.618 (1.475)
σu 47.17 KN 76.34 KN 56.67 KN max u ( t ) 224.88 KN

σxm 15.40 cm. 24.53 cm. 18.78 cm. max x m 62.44 cm.
 The rms response quantities for the stochastic res ponse analysis are presented in Table 3,
whereas the peak response quantities for the deterministic response analysis are presented in
Table 4, for selected floors and the inertial mass of damper (last row). Based on the design
code for office buildings, the maximum allowable floor acceleration is 15 cm/sec2 (or a rms
value of 5 cm/sec2). Excluding the 76 th floor on which there is no occupant, the design code
requirement is satisfied by use of an ATMD as observed from Tables 3 and 4. On the other
hand, the installation of a passive TMD does not satisfy the requirement, and the floor
accelerations are excessive for the building without any control system.

Table 3. RMS response quantities of the 76-story building for the sample controller.

No Control TMD LQG Control

σu = 56.671 KN
Floor σx io σ&x&io σx i σ&x&i σx i σ&x&i
No. 2
cm cm/s cm cm/s2 cm cm/s2
(1)
(2) (3) (4) (5) (6) (7)
1 0.016 0.256 0.009 0.255 0.008 0.240
30 2.117 2.715 1.216 2.014 1.077 1.387
50 5.145 5.111 2.953 2.766 2.617 2.027
55 6.025 5.957 3.459 3.205 3.065 2.545
60 6.936 6.674 3.982 3.343 3.529 2.624
65 7.875 7.699 4.521 4.030 4.007 3.256
70 8.829 8.473 5.070 4.187 4.493 2.997
75 9.818 9.913 5.639 5.546 4.996 3.654
76 10.040 10.343 5.766 6.020 5.109 4.945
md - - 9.655 10.492 18.778 18.120

Table 4. Peak response quantities of the 76-story building for the sample controller.

No Control TMD LQG Control

umax = 224.88 KN
Floor x pio &
x&pio x pi &
x&pi x pi &
x&pi
No.
cm cm/s2 cm cm/s2 cm cm/s2
(1)
(2) (3) (4) (5) (6) (7)
1 0.041 1.464 0.027 1.464 0.026 1.464
30 5.326 8.254 3.471 7.511 3.245 4.985
50 13.128 12.428 8.399 10.568 7.724 7.441
55 15.429 15.485 9.862 11.642 9.002 9.120
60 17.820 16.541 11.435 11.546 10.318 9.162
65 20.294 19.174 13.090 13.819 11.670 9.701
70 22.812 21.365 14.790 17.159 13.044 10.878
75 25.424 26.334 16.559 21.153 14.471 12.673
76 26.009 29.881 16.956 23.741 14.791 17.306
md - - 33.546 36.676 62.437 61.911
 The power spectral densities of the acceleration response of the 75 th floor are shown in
Fig. 5 for the building equipped with: (i) no devices, (ii) TMD, and (iii) ATMD using the
sample controller. This plot also shows the remarkable performance of the ATMD in
reducing the acceleration response.

10 -1
No Control
With TMD
ATMD (LQG)
Saa (75,75), m2 /sec3 /rad

10 -2

10 -3

10 -4

10 -5

10 -6-2
10 10 -1 10 0 10 1
Frequency (Hz)
Figure 4. Acceleration Power Spectral Densities of the 75th Floor for the Sample
Controller.

To show the robustness of the controller, we only consider the uncertainty of building
stiffness, since it has been demonstrated in the literature [e.g., Yang and Akbarpour (1990)]
that active controllers are not sensitive to the uncertainty in damping. In addition to the
building above, referred to as the "nominal building", two additional buildings are
considered; one with a 15% higher stiffness matrix and another with a 15% lower stiffness.
The ( 76 ×76) stiffness matrix (76 DOF) of the first building, referred to as the "+15%
building", is obtained by multiplying 1.15 to the ( 76 ×76) stiffness matrix of the nominal
building. Similarly, the ( 76 ×76) stiffness matrix of the second building, referred to as the
"-15% building", is obtained by multiplying 0.85 to the ( 76 ×76) stiffness matrix of the
nominal building. The controller obtained previously for the nominal building is applied to
the ± 15% buildings and the stochastic response analyses were carried out. The rms
response quantities for these two buildings with and without control are presented in Table 5.
The performance indices J1 to J 6 , which are referred to (or normalized by) the
uncontrolled response quantities of the respective buildings ( ± 15% buildings) in Table 5
are shown in Columns (2) and (3) of Table 2 for comparison.
A comparison for the results in Table 3 (nominal building) and Table 5 indicate that the
acceleration response quantities are robust. However, the displacement response and the
required actuator capacity (stroke and control force) are sensitive to the uncertainty in
stiffness. In comparison with the closed-loop response of the nominal structure (Table 3),
the displacement response, stroke and active control force for the -15% building increase by
about 24%, 30% and 35%, respectively. This is because the first natural frequency of the -
15% building is closer to that of the wind excitation and hence the uncontrolled
displacement response is bigger (Table 5). On the other hand, the displacement response,
stroke and active control force reduce by 14%, 18% and 17%, respectively, for the +15%
building in comparison with that of the nominal building (LQG controller case). This has
also been expected because the first natural frequency of the +15% building is farther away
from the major frequency of the wind excitation. Numerical results above indicate that a
designer should not overestimate the stiffness of the building in designing the controller.
The performance indices shown in Columns (2)-(4) of Table 2 indicate that they are
sensitive to the stiffness variation. However, the sensitivity is due to the wind loads rather
than the controller itself as explained previously. Thus, the robustness of the LQG controller
in this case seems to be acceptable.

Table 5. RMS Response Quantities of the 76-Story Building for the Sample Controller Using the
Building with Uncertainty in Stiffness Matrix.

Uncertainty in Stiffness (∆K) = +15 % Uncertainty in Stiffness (∆K) = -15 %

Uncontrolled LQG Control Uncontrolled LQG Control
σu = 47.166 KN σu = 76.329 KN
Floor σxio σ&x&io σx i σ&x&i σx io σ&x&io σx i σ&x&i
No.
cm cm/s2 cm cm/s2 cm cm/s2 cm cm/s2
(1)
(2) (3) (3) (4) (5) (6) (7) (8)
1 0.0110 0.2561 0.007 0.240 0.018 0.256 0.010 0.244
30 1.4269 2.3778 0.924 1.423 2.349 2.597 1.339 1.440
50 3.4661 3.9475 2.244 2.046 5.710 4.773 3.258 2.161
55 4.0589 4.5864 2.627 2.554 6.688 5.559 3.817 2.714
60 4.6722 5.0335 3.024 2.616 7.699 6.206 4.396 2.897
65 5.3044 5.8763 3.432 3.258 8.741 7.176 4.994 3.486
70 5.9467 6.3729 3.847 3.039 9.800 7.881 5.602 3.173
75 6.6128 7.7307 4.278 3.699 10.898 9.290 6.233 3.868
76 6.7620 8.1952 4.374 4.917 11.144 9.688 6.374 5.493
md ----- ----- 15.397 16.030 ----- ----- 24.537 21.430

7. CLOSURE

The 76-state building model (76 DOF), 48-state evaluation model (24 DOF with W24), the
12-state control design model, a set of 23 sample functions (time histories) of wind loads,
and the MATLAB m-files (computer codes) to; (i) execute the sample control design, and
(ii) compute the stochastic and deterministic responses of the evaluation model based on any
controller, are all available at the World Wide Website [Yang et al (1997b)]. The 46-state
model (23 DOF with W23) has also been provided for the computation of the response of the
uncontrolled building. Likewise, computer codes for the model reduction from 77 DOF
(with any ATMD and any percentage of stiffness uncertainty) to the 24 DOF with W24
(evaluation model) are included. Such computer codes will enable the designer to design
different ATMD and conduct robustness evaluations.
Passive and semi-active control systems (or devices) can be installed in any selected story
units. This will add the additional damping and stiffness to the ( 76 ×76) damping and
stiffness matrices of the nominal building. Then, the computer codes in Yang et al (1997b)
can be used to reduce the system to 23 DOF with W23 (evaluation model) and to perform
the stochastic and deterministic analyses for the determination of the response quantities of
the controlled structure. Any participant, who cannot access the World Wide Website [Yang
et al (1997b)] or has any questions regarding the benchmark problem, can contact the senior
author via E-mail at [email protected].
This is the first generation benchmark problem for the response control of wind-excited
tall buildings. Because of cumbersome computational efforts involved, many simplifications
for the control environments have been made. In particular, the actuator dynamics and
controller-structure interaction are not considered in the sample problem, although these
effects may be significant [Dyke et al (1995)]. Based on our experience, the size of the
evaluation model, i.e., 24 DOF with W24, can still be made smaller, such as 10 DOF with
W10, without compromising the accuracy. Likewise, the dimension of the controlled output
vector z can be smaller. As more experience is gain through the present benchmark study,
an improved second generation benchmark problem should be considered.

ACKNOWLEDGEMENTS

This research is partially supported by National Science Foundation Grant Nos. CMS 96-
25616 and CMS 96-15731. Valuable comments and suggestions by Professor Bill Spencer,
Jr. of Notre Dame University and the working group on benchmark problems of 2 nd
International Workshop on Structural Control led by Dr. Jay-Chung Chen of Hong Kong
University of Science and Technology are gratefully acknowledged.

APPENDIX I: STOCHASTIC RESPONSE ANALYSIS

The controller given by Eqs. (35) and (36) can be expressed in the differential form as
x&c = A c x c + B c y r (I-1)
u = Cc x c + D c y r (I-2)
in which x c = x̂ r and y r is given by Eq.(26). Since y r is a subset of y given by Eq.(3), it
can be derived directly from Eq.(3) as follows
~ ~ ~
y r = C y x + D y u + Fy W + v r (I-3)
~ ~ ~
where C y , D y and Fy are obtained from C y , D y and Fy , respectively, by condensation.
Then, the augmented closed-loop system of the evaluation model is obtained by casting
Eqs.(1), (I-1), (I-2) and (I-3) as follows
& = (A +B K ) Y + (E +B K ) W
Y (I-4)
Y Y Y Y Y w1 1
in which
 x   A 0   B   E 0 2 nxm  W 
Y =   ; AY =  ~  ; BY =  ~  ; E Y =  ~  ; W1 =  
x c  B c C y A c  B c D y  B c Fy Bc  v r 
K Y = [(I − D c D y ) − 1D c C y , (I − D c D y ) − 1C c ] ; K w1 = (I − D c D y ) − 1D c [Fy , I m ] (I-5)
~ ~ ~ ~ ~

In Eq.(I-5), 0 2nxm is a ( 2 n ×m ) zero matrix, where 2n is the dimension of the system

matrix A in the evaluation model, Eq.(1), and m is the dimension of the measurement output
vector y r (e.g., 2n = 48 and m =3 in the sample problem), and I m is an identity matrix of
dimension m. The frequency response function from W1 to Y can be obtained as
 H xW1 
H Y W1 = ( jω I − A Y − B Y K Y ) − 1 (E Y + B Y K w1 ) =   (I-6)
H xcW1 
 
in which j = ( − 1)1 / 2 . The frequency response functions from W1 to u and z are obtained
as
H uW1 = K Y H YW1 + K w1 (I-7)
and
H z W1 = C z H xW1 + D z H uW1 + [
Fz , 0 30xm ] (I-8)
where 0 30xm is a ( 30 ×m) zero matrix, since the dimension of z is 30. Hence, the rms of
z i (i-th component of z) and u can be computed by
1/ 2 1/ 2
 ∞   ∞ 
σzi =  ∫− ∞ S zizi ( ω ) dω  ; σu =  ∫− ∞ S ui u i ( ω ) dω  (I-9)
   
in which S z i z i ( ω ) and S ui ui (ω ) are the (i, i) elements of S zz (ω ) and S uu (ω ) , which are
the power spectral density matrices of z and u, respectively. Both S zz (ω ) and S uu (ω ) are
given by
S zz (ω ) = H zW1 S W1W1 H ′ zW1 ; S uu (ω ) = H uW1 S W1W1 H ′ uW1 (I-10)
In Eq. (I-9), since W and vr are uncorrelated, the power spectral density matrix S W1W1 is
given by
S ww 0 
S W1W1 =   (I-11)
 0 S vr vr 
 
Note that the acceleration response has been included in the elements of the output z,
which can be obtained from Eq. (I-9).

REFERENCES

Ankireddi, S. and Yang, H.T.Y. (1996) Simple ATMD Control Methodology for Tall Buildings
Subject to Wind Loads. ASCE Journal of Structural Engineering, Vol. 122, No. 1, pp. 83-
91.
Cao, H., Reinhorn, A.M. and Soong, T.T. (1997) Design of an Active Mass Damper for Wind
Response of Nanjing TV Tower. to appear in Journal of Engineering Structures.
Davison, E.J. (1966) A Method for Simplifying Linear Dynamic Systems. IEEE Transactions on
Automatic Control, Vol. AC-11, No. 1, pp. 93-101.
Dyke, S.J., Spencer, B.F. Jr., Quast, P. and Sain, M.K. (1995 ) The Role of Control Structure
Interaction in Protective System Design. Journal of Engineering Mechanics, ASCE, Vol.
121, No. 2, pp. 322-338, 1995.
Housner, G.W. and Masri, S.F., Eds. (1994). Proceedings of the First World Conference on Structural
Control, International Association for Structural Control, USC publication, Los Angeles.
Kareem, A. (1981) Wind-excited Response of Buildings in Higher Modes. ASCE Journal of
Structural Division, Vol. 107, No. ST4, pp.701-706.
Samali, B., Yang J. N. and Yeh, C. T. (1985 ) Control of Lateral-Torsional Motion of Wind-Excited
Buildings. ASCE Journal of Engineering Mechanics, Vol. 111, No. 6, pp.777-796.
Shinozuka, M. (1985) Stochastic Fields and Their Digital Simulation. Lecture Notes for CISM Course
on Stochastic Methods in Structural Mechanics, Hermes Publication Ltd, Udine, Italy,
August 28-30.
Simiu, E. and Scanlan, R.H. (1986) Wind Effects on Structures, 2nd Ed., John Wiley & Sons Inc.
Skelton, R.E. (1988) Dynamic Systems Control: Linear Systems Analysis and Synthesis, Wiley, New
York.
Spencer, Jr., B.F., Dyke, S.J., and Deoskar, H.S. (1998) Benchmark Problems in Structural Control-
Part 1: Active Mass Driver System, and Part 2: Active Tendon System. Paper to appear in
Journal of Earthquake Engineering and Structural Dynamics.
Suhardjo, J., Spencer, B.F. and Kareem, A (1992 ) Frequency Domain Optimal Control of Wind-
Excited buildings. ASCE Journal of Engineering Mechanics, Vol. 118, No. 12, pp.2463-
2481.
Wu J. C. and Yang, J. N. (1997a ) Continuous Sliding Mode Control of a TV Transmission Tower
Under Stochastic Wind. Proceedings of 1997 American Control Conference , Vol 2, pp.883-
887.
Wu, J. C. and Yang, J.N. (1997b ) LQG and H-Infinity Control of a TV Transmission Tower Under
Stochastic Winds. Proceedings of the International Conference on Structural Safety and
Reliability, Nov. 24-28, 1997, Kyoto, Japan.
Wu, J. C., Yang, J. N. and Schmitendorf, W. (1998 ) Reduced-order H-Infinity and LQR Control for
Wind-Excited Tall Buildings. to appear in Journal of Engineering Structures.
Yang, J. N. and Samali, B. (1983 ) Control of Tall Buildings in Along-Wind Motion. ASCE Journal of
Structural Engineering, Vol. 109, No.1, pp. 50-68.
Yang, J.N. and Akbarpour, A. (1990) Effect of System Uncertainty on Control of Seismic-Excited
Buildings. Journal of Engineering Mechanics, ASCE, Vol. 116, No. 2, pp. 462-478.
Yang, J. N., Wu, J.C., Agrawal A. K. and Hsu, S. Y. (1997a ) Sliding Mode Control with
Compensators for Wind and Seismic Response Control. Journal of Earthquake Engineering
and Structural Dynamics, Vol. 26, pp. 1137-1156.
Yang, J.N., Wu, J.C., Samali, B. and Agrawal, A.K. (1997b ) A Benchmark Problem for Response
Control of Wind-Excited Tall Buildings. Benchmark Problem Package Available at the
World Wide Website: http://www.eng.uci.edu/civil/civileng.html.