Engineering Structures 255 (2022) 113902

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Buckling in prestressed stayed beam–columns and intelligent evaluation

Kaidong Wu a , Xuhong Qiang b ,∗, Zhe Xing a,c , Xu Jiang b
a
College of Civil and Transportation Engineering, Hohai University, Nanjing 210098, China
b
College of Civil Engineering, Tongji University, Shanghai 200092, China
c
Zhejiang Southeast Space Frame Co., Ltd., Hangzhou 311209, China

ARTICLE INFO ABSTRACT

Keywords: For prestressed stayed beam–columns, the buckling behaviour is studied analytically, and the intelligent models
Prestressed stayed beam–columns for evaluating the nonlinear behaviour are developed using machine learning. An implicit solution, which
Steel structures allows the direct evaluation for the buckling behaviour of both doubly-symmetric and mono-symmetric stayed
Interactive buckling
members, is determined analytically for the first time, and an explicit simplified solution of the buckling load
Analytical and finite element modelling
is obtained using regression. Both the implicit and explicit solutions can simplify the traditional numerical
Machine learning
method for evaluating the buckling load of stayed members, and they are verified using numerical modelling
and excellent comparisons are obtained. Based on the explicit solution of the buckling load, machine learning
models are adopted to develop the intelligent methods for evaluating the nonlinear buckling behaviour of
prestressed stayed beam–columns, which can overcome the difficulties in distinguishing between pure and
interactive buckling modes in the traditional method for load-carrying capacity evaluation. The results in the
test set show that Extreme Gradient Boosting (XGBoost) and Artificial Neural Network (ANN) can predict the
nonlinear failure mode and ultimate load of prestressed stayed beam–columns accurately and reliably.

1. Introduction Apart from the vertical stayed column with equal crossarms, the
crossarm lengths can be different to optimize the structural behaviour
Compared to unstayed members, prestressed stayed members have under combined axial and bending loads [33–36], so-called ‘mono-
more excellent mechanical performance owing to cable-stayed effects, symmetric stayed beam–column’, as illustrated in Fig. 1. An iconic
since they can maintain load-carrying capacity with lighter self-weight. structure that has already employed a system of stayed columns and
The crossarm-cable system can provide additional supports between beam–columns is ‘La Grande Arche’ in Paris [37]. The internal structure
two member ends, which can reduce the effective length of the main includes a suspended membrane roof, part of which is supported by
member and increase the buckling load significantly. The investiga- a system of prestressed cable-stayed columns that are braced every
tion on the stayed columns started in 1960s [1–3], and the early 21 m by stayed beam–columns in the horizontal plane, the details of
studies were on the buckling load. Finite element (FE) modelling was which are shown in Fig. 2. The current work continues the studies on
applied firstly in [4,5]. These aforementioned works on the buck- the stayed beam–columns. Since the buckling eigenvalues prior from
ling behaviour were validated further using testing in [6–9], and the
ABAQUS [38] is required to evaluate the linear optimum prestressing
stability of the stayed column with multiple crossarms was investi-
force in [33–35], it is not convenient to develop a design method.
gated in [10–12]. Moreover, the relationship between the post-buckling
Moreover, although there exist the studies on the design method for the
behaviour and the prestressing level was determined [13–15], and
stayed member [20,21], they just focus on the vertical stayed columns,
the interactive buckling was observed and studied [16–19]. The de-
and their accuracy is not satisfied. Currently, the artificial intelligence
sign and optimization method for the stayed column were proposed
in [20–23]. In addition, the stability and mechanical performance of (AI) technology provides a new way to solve the aforementioned prob-
prestressed stayed members made of different metals were investi- lems [39–42]. The current work is to investigate the buckling behaviour
gated [24,25]. The mechanical performance of the stayed elements of the prestressed stayed beam–columns analytically and propose an
under other loading cases were studied in [26–29]. Recently, the intelligent method for evaluating the nonlinear failure mode and ulti-
mechanical behaviour of prestressed stayed FRP reinforcement steel mate load using machine learning. An analytical solution to determine
members was investigated [30–32]. the buckling load of stayed members, which can be applied to both the

∗ Corresponding author.
E-mail addresses: [email protected] (K. Wu), [email protected] (X. Qiang), [email protected] (Z. Xing), [email protected] (X. Jiang).

https://doi.org/10.1016/j.engstruct.2022.113902
Received 24 September 2021; Received in revised form 6 January 2022; Accepted 13 January 2022
Available online 2 February 2022
0141-0296/© 2022 Elsevier Ltd. All rights reserved.
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 1. Prestressed stayed beam–column: 𝑃 is the axial load; 𝑄a and 𝑞w are the crossarm and main member self-weight respectively; 𝑇 is the prestressing force within stays.

Fig. 2. Application of prestressed stayed columns and beam–columns within La Grande Arche in Paris: (a) structural elevation; (b) detail from the circled region in (a) showing
stayed elements in the vertical and horizontal plane.
Source: Schematics have been taken from [37].

doubly-symmetric and mono-symmetric arrangements, is obtained. In crossarms, the upper crossarm length ratio 𝛼u , the lower crossarm
addition, Extreme Gradient Boosting (XGBoost) and Artificial Neural length ratio 𝛼l ,and crossarm length ratio 𝛼 can be defined thus:
Network (ANN) are adopted to evaluate both the nonlinear failure 𝑎 𝑎l 𝛼
mode and ultimate load. The studies on the analytical buckling load 𝛼u = u , 𝛼l = , 𝛼 = u. (1)
𝐿 𝐿 𝛼l
and intelligent method for evaluating the failure mode and ultimate √
load of the stayed beam–columns would be developed to an efficient The radius of gyration for the main member is 𝑟c = 𝐼c ∕𝐴c . For the
design method for this type structure independent to the finite element main member, 𝐸c , 𝐴c and 𝐼c represent its Young’s modulus, cross-
analysis. sectional area and second moment of area respectively. For the upper
and lower crossarms, they have the same material and cross-sectional
2. General definitions and modelling properties to these of the main member, i.e. 𝐸a = 𝐸c , 𝐴a = 𝐴c and
𝐼a = 𝐼c . The Young’s modulus and cross-sectional area of all the stays
In the current section, the studied parameters are defined herein, are same, which are expressed as 𝐸s and 𝐴s respectively. Based on these
including the structural parameters and benchmark prestressing levels, basic parameters, the axial stiffnesses of the main member 𝐾c,a , upper
and the FE modelling is presented. crossarm 𝐾au,a , lower crossarm 𝐾al,a , upper stay 𝐾su and lower stay 𝐾sl
can be obtained thus:
2.1. Structural parameters 𝐸𝐴 𝐸𝐴 𝐸𝐴
𝐾c,a = c c , 𝐾au,a = a a , 𝐾al,a = a a ,
𝐿 𝑎u 𝑎l
The definitions of the stays and the studied geometric parameters 𝐸s 𝐴s 𝐸s 𝐴s
𝐾su = , 𝐾sl = . (2)
are shown in Fig. 3. Based on the lengths of the main member and 𝑙u 𝑙l

2
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 3. Definitions for the stays and geometric parameters of the stayed beam–column: 𝐿 is the main member length; 𝑎u and 𝑎l are the upper and lower crossarm lengths
respectively; 𝛽u and 𝛽l are the stay angles with respect to the main member; 𝑙un and 𝑙ln are the upper and lower stay lengths respectively, where n = {1, 2}. Currently, 𝑙u = 𝑙u1 = 𝑙u2
and 𝑙l = 𝑙l1 = 𝑙l2 .

Table 1 Table 3
Constant structural parameters in the current study. Imperfection combinations: cases 1 and 3 are the pure eigenmode imperfections for
Length of main member 𝐿 = 3050 mm half-sine symmetric mode, and the negative and positive are corresponding to the
Main member and crossarm Young’s moduli 𝐸a = 𝐸c = 201 kN∕mm2 downward and upward imperfections respectively; case 2 is to trigger the interactive
Stay Young’s moduli 𝐸s = 202 kN∕mm2 buckling [34,35].
Outside diameter of main member 𝜙co = 38.1 mm Case 1 2 3
Inside diameter of main member 𝜙ci = 25.4 mm
𝜇1 −1.000 −0.750 1.000
Outside diameter of crossarm 𝜙ao = 38.1 mm
𝜇2 0.000 −0.331 0.000
Inside diameter of crossarm 𝜙ai = 25.4 mm
Material density (steel) 𝜌 = 7850 kg∕m3
Gravitational acceleration 𝑔 = 9.81 m∕s2
Yield stress of main member 𝑓y = 340 N∕mm2 𝑇opt2 sin 𝛽u
𝑇opt21 = , (4)
sin 𝛽l
Table 2 C C
Parameter values of the doubly-symmetric (𝛼 = 1.0) cases, where 𝛼l = 0.1 in cases F1–F6
where the maximum critical buckling load 𝑃max = 𝑃FE,𝑇 =0
∕𝐶0 with
C
𝑃FE,𝑇 defined as the elastic buckling load of a stayed column with
and 𝜙s ∕(2𝑟c ) = 0.21 in cases A1–A6. =0
Case F1 F2 F3 F4 F5 F6 the zero prestressing force, which is determined using ABAQUS. 𝐶0 , 𝐶u
𝜙s (mm) 1.6 3.2 4.8 6.4 8.0 10.0 and 𝐶l are the quantities from the analytical works in [33].
𝜙s ∕(2𝑟c ) 0.07 0.14 0.21 0.28 0.35 0.44
Case A1 A2 A3 A4 A5 A6
2.3. Finite element (FE) modelling
𝛼l 0.025 0.05 0.075 0.100 0.125 0.150

Following the previous FE studies on the stayed members [35,43,

44], the commercial software ABAQUS [38] is used for the current
The bending stiffnesses of the main member 𝐾c,b , upper crossarm 𝐾au,b numerical work. The Euler–Bernoulli beam element B23 is adopted for
and lower crossarm 𝐾al,b can be obtained thus: the main member and the crossarms, and the element length of them
𝐸c 𝐼c 𝐸a 𝐼a 𝐸a 𝐼a is 50 mm. The truss element T2D2 with ‘no compression’ option is
𝐾c,b = , 𝐾au,b = , 𝐾al,b = . (3)
𝐿3 𝑎3u 𝑎3l adopted for the cable, and each cable is modelled as an individual
element. The linear buckling load and buckling profile are obtained
Apart from 𝑎l , 𝛼 and stay diameter 𝜙s , other structural parame-
using the Buckling Analysis in ABAQUS. From [35], the combined
ters are constants, and their values are listed in Table 1. In terms
imperfection shape for the main member 𝑊0 is introduced to trigger
of doubly-symmetric cases, 12 configurations, i.e. F1–F6 and A1–
different buckling modes such as pure mode buckling and interactive
A6 from [16], are studied. In terms of the mono-symmetric cases,
9 mono-symmetric cases can be generated from 1 doubly-symmetric buckling, thus:
case by varying the crossarm length ratio 𝛼 from 0.1 to 0.9. For [ ]
𝜋𝑥 2𝜋𝑥
𝑊0 = 𝛿𝐿 𝜇1 sin + 𝜇2 sin , (5)
example, mono-symmetric cases F3.1–F3.9 are corresponding to 𝛼 = 𝐿 𝐿
{0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9} respectively, which are generated where 𝛿 is the imperfection amplitude ratio; 𝜇1 and 𝜇2 are the scale
from the doubly-symmetric case F3. The detailed parameters of each factors to indicate the proportions of the initial imperfection shape. The
studied doubly-symmetric case are listed in Tables 2. Therefore, 120
relationship between 𝜇1 and 𝜇2 is:
cases including 12 doubly-symmetric cases and 108 mono-symmetric
cases are investigated in the current work. Varied 𝜙s ∕(2𝑟c ) indicates 𝜇12 + 4𝜇22 = 1. (6)
the varied relative stiffness between the cables and the main member;
varied 𝛼l , 𝛼u and 𝛼 allow the stay angles to be varied. The studied imperfection combinations in the current work are listed
in Table 3.
2.2. Benchmark prestressing force

For the doubly-symmetric (𝛼 = 1.0) stayed column, the benchmark 3. Analytical solutions of linear buckling load
prestressing forces and the relationship between the critical buckling
load and the prestressing level were proposed in [3,14], and they have The principal purpose of the current section is to obtain the ana-
been developed further for the mono-symmetric (𝛼 < 1.0) stayed beam– lytical solutions of the linear buckling load for the prestressed stayed
column in [33]. From [33], the optimum prestressing forces 𝑇opt1 , 𝑇opt2 , beam–columns. These solutions can be used as 𝑃FE,𝑇 C in Eq. (4).
=0
𝑇opt12 and 𝑇opt21 can be determined as follows: Moreover, they are the basis to develop the intelligent methods for
𝑇opt1 sin 𝛽l evaluating the nonlinear failure mode and ultimate load of the stayed
C C
𝑇opt1 = 𝑃max 𝐶l , 𝑇opt2 = 𝑃max 𝐶u , 𝑇opt12 = , beam–columns in the next section.
sin 𝛽u

3
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 4. Illustration of buckling eigenmodes 1 with half-sine symmetric deflection and 2 with full-sine antisymmetric deflection. Currently, they are defined as Mode 1 and Mode
2 respectively.

Fig. 5. Model and free body diagram for Mode 1 buckling; 𝑄1 is the vertical force on the main member ends resulting from Mode 1 buckling; 2𝑄1 at the main member mid-span
is the total vertical force from the stay restraints, currently modelled as a linear longitudinal spring; 𝑙 is defined as the half-length the main member 𝐿∕2; 𝑁, 𝑆 and 𝑀 are the
cross-sectional axial force, shear force and moment respectively.

3.1. Assumptions where the stayed beam–column becomes unstable, it is necessary to

consider force equilibrium of the main member in its deflected shape
In a similar method to the analytical work in [2], for a mono- using the free body diagram in Fig. 5. The bending moment at any
symmetric stayed member, two different buckling modes, Modes 1 cross-section 𝑀 can be expressed thus:
and 2 are defined, as illustrated in Fig. 4. The lowest buckling load
between these two modes is the critical buckling load. The following 𝑀 = 𝑃 𝑤 − 𝑄1 𝑥. (8)
assumptions are also made for the analytical model presently: (a) Lin- Substituting Eq. (8) into Eq. (7) gives:
early elastic material; (b) Small deflections; (c) No second-order terms;
(d) Perfect geometry; (e) Buckling in plane; (f) Zone 3 prestressing d2 𝑤 d2 𝑤 𝑄1
𝐸c 𝐼c + 𝑃 𝑤 = 𝑄1 𝑥 ⟹ + 𝑘2 𝑤 = 𝑥, (9)
d𝑥 2 d𝑥2 𝐸c 𝐼c
level [33].
√
where 𝑘 = 𝑃 ∕(𝐸c 𝐼c ). Eq. (9) is the governing differential equation for
3.2. Theoretical solutions the Mode 1 buckling. The general solution of Eq. (9) is:
𝑄1
From all the aforementioned assumptions, the buckling load for 𝑤 = 𝐴1 sin 𝑘𝑥 + 𝐴2 cos 𝑘𝑥 + 𝑥, (10)
each mode can be determined by solving beam differential equations, 𝑃
and the basic differential equation for beam bending is: where 𝐴1 and 𝐴2 are constants that can be determined from the Mode
1 geometric boundary conditions, as shown in Fig. 6. Three conditions
d2 𝑤 (7)
𝑀 = −𝐸c 𝐼c , are required to obtain the particular solution: (a) At 𝑥 = 0, 𝑤 = 0; (b)
d𝑥2
At 𝑥 = 𝑙, 𝑤 = 𝑤m ; (c) At 𝑥 = 𝑙, d𝑤d𝑥
= 0. The final condition is one of
where 𝑥 and 𝑤 are the main member axis (horizontal) and lateral reflective symmetry that is specific to Mode 1. Applying these boundary
(vertical) deflection function respectively; 𝑀 is the bending moment conditions leads to the following solution for the mid-span deflection
at any cross-section. The theoretical buckling solution for each mode is
𝑤 = 𝑤m :
formulated subsequently.
𝑄1
𝑤m = (𝑘𝑙 − tan 𝑘𝑙). (11)
3.2.1. Mode 1 buckling 𝑃𝑘
The Mode 1 buckling shape of the stayed beam–column can be To complete the buckling solutions, it is necessary to determine the
replaced by the model as shown in Fig. 5. To establish the condition relationship between 𝑄1 and 𝑤m . The vertical equilibrium of the forces

4
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 6. Forces and deformations resulting from Mode 1 buckling; 𝑤m is the mid-span vertical displacement of the main member; 𝐹u,f and 𝐹l,f are the forces within the upper and
lower crossarms respectively at the instant of buckling; 𝛥𝐹u,f and 𝛥𝐹l,f are the absolute changes of force within the upper and lower crossarms respectively resulting from the
beam–column buckling.

acting on the main member at the crossarms (see Fig. 6) requires that: 3.2.2. Mode 2 buckling
The Mode 2 buckling shape of the stayed beam–column can be
replaced by the model shown in Fig. 7, For this case, the bending
2𝑄1 = (𝐹l,f + 𝛥𝐹l,f ) − (𝐹u,f − 𝛥𝐹u,f ). (12)
moment at any cross-section can be expressed thus:
From the relationship of the prestressing forces within the upper and
𝑀 = 𝑃 𝑤 − 𝑄2 (𝑙 − 𝑥). (19)
lower stays, 𝑇u,i sin 𝛽u = 𝑇l,i sin 𝛽l presented in [33] is introduced, and
the axial forces within the crossarms are equal, thus 𝐹u,f = 𝐹l,f , which Substituting Eq. (19) into Eq. (7) leads to:
gives:
d2 𝑤 𝑄2
+ 𝑘2 𝑤 = (𝑙 − 𝑥), (20)
2𝑄1 = 𝛥𝐹u,f + 𝛥𝐹l,f = 2𝛥𝑇u,f sin 𝛽u + 2𝛥𝑇l,f sin 𝛽l , (13) d𝑥2 𝐸c 𝐼c
which represents the governing differential equation for Mode 2 buck-
where 𝛥𝑇u,f and 𝛥𝑇l,f are the absolute changes of forces within the
ling. The general solution of Eq. (20) is:
upper and lower stays respectively resulting from the beam–column
buckling. The absolute changes in length of the upper and lower stays, 𝑄2
𝑤 = 𝐴3 sin 𝑘𝑥 + 𝐴4 cos 𝑘𝑥 + (𝑙 − 𝑥), (21)
𝛥𝑙u,m1 and 𝛥𝑙l,m1 , can be expressed respectively as shown in Fig. 6 thus: 𝑃
where 𝐴3 and 𝐴4 are constants of integration that can be determined
from the Mode 2 geometric boundary conditions, as shown in Fig. 8;
𝛥𝑙u,m1 = 𝑤m sin 𝛽u , 𝛥𝑙l,m1 = 𝑤m sin 𝛽l . (14)
deflections 𝑤au,r , 𝑤au,b , 𝑤au,f , and forces 𝑉u,f and 𝛥𝑉u,f are for the upper
This leads to the expressions of the absolute changes of forces within stays; deflections 𝑤al,r , 𝑤al,b , 𝑤al,f , and forces 𝑉l,f and 𝛥𝑉l,f are for
the upper and lower stays, 𝛥𝑇u,f and 𝛥𝑇l,f respectively, thus: the lower stays. Three conditions are required to obtain the particular
solution: (a) At 𝑥 = 0, 𝑤 = 0; (b) At 𝑥 = 𝑙, 𝑤 = 0; (c) At 𝑥 = 0,
𝛥𝑇u,f = 𝐾su 𝛥𝑙u,m1 = 𝐾su 𝑤m sin 𝛽u , 𝑑𝑤
= 𝜃cm . The final condition being the one for rotational symmetry
(15) 𝑑𝑥
𝛥𝑇l,f = 𝐾sl 𝛥𝑙l,m1 = 𝐾sl 𝑤m sin 𝛽l , that is specific to Mode 2. Applying these boundary conditions leads to
the following equation:
where 𝐾su and 𝐾sl are the stiffness of upper and lower stays defined in
Eq. (2). Substituting Eq. (15) into Eq. (13) gives: 𝑘𝑙
tan 𝑘𝑙 = 𝑃 𝜃cm
. (22)
1+
𝑄1 = 𝐾su 𝑤m sin2 𝛽u + 𝐾sl 𝑤m sin2 𝛽l . (16) 𝑄2

To determine an appropriate solution for Mode 2 buckling, it is nec-

Substituting Eq. (16) into Eq. (11) and rearranging gives:
essary to determine the relationship between 𝜃cm and 𝑄2 . From Fig. 7,
(𝑘𝑙)3 moment equilibrium acting on the free body of the main member in
= 𝐷1 , (17)
𝑘𝑙 − tan 𝑘𝑙 conjunction with the crossarms requires that the following expression
where the new quantity 𝐷1 is: is satisfied:

𝐾su sin2 𝛽u + 𝐾sl sin2 𝛽l 2𝑄2 𝑙 = 𝑅au 𝑎u + 𝑅al 𝑎l , (23)

𝐷1 = , (18)
𝐾cb,half
where 𝑎u = 𝑙 tan 𝛽u and 𝑎l = 𝑙 tan 𝛽l from the preceding geometric
and the bending stiffness of the half main member is 𝐾cb,half = 𝐸c 𝐼c ∕𝑙3 . definitions. From the equilibrium of forces on the tip of the crossarm

5
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 7. Model and free body diagram for Mode 2 buckling; 𝑄2 is the vertical force on the main member ends resulting from Mode 2 buckling; 𝑅au and 𝑅al are the total horizontal
forces on the upper and lower crossarm tips from the upper and lower stay restraints respectively; 𝑁, 𝑆 and 𝑀 are the cross-sectional axial force, shear force and moment
respectively.

Fig. 8. Forces, displacements and rotations resulting from Mode 2 buckling; 𝜃cm and 𝜃a are the rotational angles of the main member and the upper crossarm at their intersection
point respectively; 𝑤al,r is the rigid body displacement of the lower crossarm tip owing to the joint rotation; 𝑤al,b is the displacement of the lower crossarm tip owing to the
crossarm bending; 𝑤al,f is the final displacement of the lower crossarm tip; 𝑉l,f is the horizontal force on the lower crossarm tip at the instant of buckling; 𝛥𝑉l,f is the absolute
change of the horizontal force on the lower crossarm tip resulting from buckling.

6
K. Wu et al. Engineering Structures 255 (2022) 113902

shown in Fig. 8, the following equations are obtained: the axial force within the main member arising from the stay stress can
be expressed as:
𝑅au = 2𝛥𝑉u,f = 2𝛥𝑇u,f cos 𝛽u , 𝑅al = 2𝛥𝑉l,f = 2𝛥𝑇l,f cos 𝛽l . (24) ( )
sin 𝛽u
Owing to the small lateral displacement of the main member at instabil- 𝑇l,i cos 𝛽l + 𝑇u,i cos 𝛽u = cos 𝛽l + cos 𝛽u × 𝑇u,i
sin 𝛽l
ity, the absolute changes in length of the upper and lower stays, 𝛥𝑙u,m2 ( )
sin 𝛽u
and 𝛥𝑙l,m2 , can be expressed respectively, as shown in Fig. 8, thus: ≤ cos 𝛽l + cos 𝛽u 𝑓y 𝐴s , (35)
sin 𝛽l
𝛥𝑙u,m2 = 𝑤au,f cos 𝛽u , 𝛥𝑙l,m2 = 𝑤al,f cos 𝛽l , (25) where 𝑓y is the yielding stress of applied material. Considering the
status where there is no external axial load, the stress within stays may
which leads to the expressions for the absolute changes of the forces
lead to the buckling of the main member, and the following inequality
within the upper and lower stays, 𝛥𝑇u,f and 𝛥𝑇l,f respectively, thus:
needs to be satisfied, thus:
𝛥𝑇u,f = 𝐾su 𝛥𝑙u,m2 = 𝐾su 𝑤au,f cos 𝛽u , 𝜋 2 𝐸c 𝐼c
(26) 𝑇l,i cos 𝛽l + 𝑇u,i cos 𝛽u ≤ 𝑃 C → 𝑃E = , (36)
𝛥𝑇l,f = 𝐾sl 𝛥𝑙l,m2 = 𝐾sl 𝑤al,f cos 𝛽l . 𝑙e2
Substituting Eqs. (24)–(26) into Eq. (23) gives: where 𝑙e = 𝑛𝑙 is the effective length, and 𝑛 is the effective length factor
for one-half of the main member. To maximize the material efficiency
𝑄2 = 𝐾su 𝑤au,f cos 𝛽u sin 𝛽u + 𝐾sl 𝑤al,f cos 𝛽l sin 𝛽l . (27)
of both the stays and main member, from Eqs. (35) and (36), the
From the displacement definitions in Fig. 8, the relationships among following equation is obtained:
different displacements of the upper and lower crossarm tips are as ( )
𝜋 2 𝐸c 𝐼c sin 𝛽u
follows: = cos 𝛽l + cos 𝛽u 𝑓y 𝐴s . (37)
(𝑛𝑙)2 sin 𝛽l
𝑤au,r = 𝑤au,f + 𝑤au,b = 𝜃a 𝑎u , 𝑤al,r = 𝑤al,f + 𝑤al,b = 𝜃a 𝑎l . (28) Replacing 𝑓y = 𝐸s 𝜀s,y where 𝜀s,y is the yield strain of the stay material
Based on Euler–Bernoulli bending theory, the bending displacements of and rearranging Eq. (37) gives:
the upper and lower crossarm tips, 𝑤au,b and 𝑤al,b , are given thus: 𝐴s 𝑙2 1 𝜋2 𝛺
= = , (38)
𝑅au 2𝐾su 𝐼c 𝑛2 𝜀s,y sin 𝛽u
cos 𝛽l + cos 𝛽u 𝑛2 𝜀s,y
𝑤au,b = = 𝑤 cos2 𝛽u , sin 𝛽l
3𝐾au,b 3𝐾au,b au,f
(29) where 𝛺 is defined thus:
𝑅al 2𝐾sl
𝑤al,b = 𝑤 cos2 𝛽l .
3𝐾al,b 3𝐾al,b al,f 𝜋2
𝛺= . (39)
sin 𝛽u
Substituting Eqs. (29) into Eq. (28) and rearranging gives: sin 𝛽l
cos 𝛽l + cos 𝛽u

𝑎u 𝜃a 𝑎l 𝜃a Since explicit simplified solutions are required such that the solutions
𝑤au,f = , 𝑤al,f = . (30)
2𝐾su 2𝐾sl can be obtained without requiring iterations, only simple functions are
1+ 3𝐾au,b
cos2 𝛽u 1+ 3𝐾al,b
cos2 𝛽l
taken into account. Nonlinear Least Squares (NLS) [45] is adopted to
Substituting Eq. (30) into Eq. (27) gives: obtain the simplified solutions of the buckling loads. The sum of the
⎡ ⎤ error squared between the true values and fitted values (represented as
𝑎u 𝜃 a
𝑄2 = 𝐾su ⎢ ⎥ cos 𝛽u sin 𝛽u 𝑅2 ) is adopted as the cost function in this optimization. From Eq. (34),
⎢1 + 2𝐾su
cos2 𝛽u ⎥⎦ it can be seen that 𝐹1,theory and 𝐹1,theory approach infinity at 𝑘𝑙 = 𝜋∕0.7.
⎣ 3𝐾au,b
(31) Moreover, since explicit simplified solutions are required such that
⎡ 𝑎l 𝜃a ⎤
+ 𝐾sl ⎢ ⎥ cos 𝛽l sin 𝛽l . the solutions can be obtained without requiring iterations, only simple
⎢1 + 2𝐾sl
cos2 𝛽l ⎥⎦ functions are taken into account. Considering the aforementioned two
⎣ 3𝐾al,b
properties for 𝐹1,theory and 𝐹2,theory , the results based on 𝐹𝑖,regression =
From the aforementioned assumptions, 𝜃cm = 𝜃a owing to the rigid joint a + b𝑒c𝑥+d and 𝐹𝑖,regression = a + b tan(c𝑥 + d) are compared, and the
at mid-span. Substituting Eq. (31) into Eq. (22) gives: latter is adopted finally owing to the excellent fit that is found. All these
(𝑘𝑙)2 tan 𝑘𝑙 computations were conducted within Matlab.
= 𝐷2 , (32)
𝑘𝑙 − tan 𝑘𝑙
where the new quantity 𝐷2 is: 3.3.1. Mode 1 buckling
From Fig. 4, for Mode 1 buckling, the effective length of half of
𝐾su ⎡⎢ sin2 𝛽u ⎤
⎥ + 𝐾sl
⎡
⎢ sin2 𝛽l ⎤
⎥. the main member in the stayed beam–column is between that of the
𝐷2 = (33)
𝐾cb,half ⎢ 1 + 2𝐾 su
cos2𝛽 ⎥ 𝐾cb,half ⎢1 + 2𝐾 sl
cos2 𝛽l ⎥⎦ cantilever and that of the case where one end is pinned and the other
⎣ 3𝐾au,b u⎦ ⎣ 3𝐾al,b
is fixed. Hence, 𝑛 = [0.7, 2], and 𝑘𝑙 = [𝜋∕2, 𝜋∕0.7] ≈ [1.571, 4.488],
3.3. Regression solutions rearranging Eq. (18) gives:
𝐴s 𝑙 2
The buckling loads for Modes 1 and 2 can be obtained by solving 𝐷1 = (cos 𝛽u sin2 𝛽u + cos 𝛽l sin2 𝛽l ). (40)
𝐼c
Eqs. (17) and (32) respectively. However, these two equations are
Substituting Eq. (38) into (40) gives:
implicit and cannot be solved in closed form and need to be simplified.
Assuming 𝐸c = 𝐸a = 𝐸s and 𝐼c = 𝐼a , the left-hand side of Eqs. (17) and 𝛺
𝐷1 = (cos 𝛽u sin2 𝛽u + cos 𝛽l sin2 𝛽l ). (41)
(32) can be defined as two functions, 𝐹1,theory and 𝐹2,theory respectively, 𝑛2 𝜀s,y
thus: From [46], for the commonly used carbon steel, the yield stress 𝑓y =
(𝑘𝑙)3 (𝑘𝑙)2 tan 𝑘𝑙 [215, 460] MPa, and Young’s Modulus 𝐸 = 210 GPa; hence, 𝜀s,y =
𝐹1,theory = , 𝐹2,theory = . (34)
𝑘𝑙 − tan 𝑘𝑙 𝑘𝑙 − tan 𝑘𝑙 [0.10%, 0.22%]. In addition, 𝛼l = [0.01, 0.15], 𝛼 = [0.1, 1.0] and 𝑛 =
From Eqs. (17), (18), (32) and (33), the values of these two functions (0.7, 2). Finally, 𝐷1 = [0.37, 1663.11] can be determined numerically, and
should be positive and approach positive infinity as 𝑘𝑙 → 𝜋∕0.7 ≈ 4.49. 𝐷1 = [0, +∞) corresponding to 𝑘𝑙 = [𝜋∕2, 𝜋∕0.7) can be adopted as the
To obtain an approximate, from [33], 𝑇l,i sin 𝛽l = 𝑇u,i sin 𝛽u is introduced, required range of 𝐹1,theory to determine 𝐹1,regression . Applying NLS, the

7
K. Wu et al. Engineering Structures 255 (2022) 113902

where 𝑖 = {1, 2} corresponds to Mode 1 or 2 respectively. Rearranging

Eq. (46) gives the regression solution of the buckling load for Mode 𝑖,
𝐶
i.e. 𝑃𝑖,regression , thus:
( 𝐷 −𝑏 ) 2
⎡ arctan 𝑖 𝑖,1
− 𝑏𝑖,4 ⎤
𝐸𝐼 ⎢ 𝑏𝑖,2 ⎥
𝐶
𝑃𝑖,regression = c c⎢ ⎥ , (47)
𝑙2 ⎢ 𝑏𝑖,3 ⎥
⎣ ⎦
where 𝑏𝑖,𝑗 is the parameter from regression, and 𝑖 = {1, 2}, 𝑗 =
{1, 2, 3, 4}. In terms of Mode 1 buckling (𝑖 = 1), from Eq. (42), 𝑏1,𝑗 can
be determined thus:
⎧
⎪𝑏1,1 = −3.4206, 𝑏1,2 = 7.0485, 𝑏1,3 = 0.4573, 𝑏1,4 = −0.2671,
⎪if 𝐷1 = (0, 4.42];
⎪
⎪𝑏1,1 = 0.2259, 𝑏1,2 = 11.2878, 𝑏1,3 = 0.4753, 𝑏1,4 = −0.7939,
⎨ (48)
⎪if 𝐷1 = [4.42, 30.18];
⎪𝑏 = 15.9457, 𝑏 = 2.3136, 𝑏 = 0.5139, 𝑏 = −0.7384,
⎪ 1,1 1,2 1,3 1,4
Fig. 9. Comparison between the theory and regression functions for the buckling loads. ⎪if 𝐷1 = [30.18, +∞).
⎩
𝐹1,theory and 𝐹2,theory are from Eq. (34), and 𝐹1,regression and 𝐹2,regression are from Eqs. (42)
and (45) respectively. The dashed line is 𝑘𝑙 ≈ 4.49, where 𝐹1,theory , 𝐹2,theory , 𝐹1,regression In terms of Mode 2 buckling (𝑖 = 2), from Eq. (45), 𝑏2,𝑗 can be
and 𝐹2,regression approximate at infinite. determined thus:

𝑏2,1 = −2.3799, 𝑏2,2 = 3.2130, 𝑏2,3 = 0.6999, 𝑏2,4 = −1.5753,

following expressions are obtained: 𝐷2 = [1.45, 25.00]. (49)

⎧ To verify the aforementioned theoretical and regression solutions

⎪−3.4206 + 7.0485 tan(0.4573𝑘𝑙 − 0.2671), of the buckling loads for Modes 1 and 2, linear buckling analysis
⎪ if 𝑘𝑙 ∈ (𝜋∕2, 2.42];
⎪ via ABAQUS is conducted. The aforementioned mono-symmetric cases
⎪0.2259 + 11.2878 tan(0.4753𝑘𝑙 − 0.7939), F3.1–F3.9 are investigated. Fig. 10 compares the buckling loads from
𝐹1,regression =⎨ (42)
⎪ if 𝑘𝑙 ∈ [2.42, 4.18]; FE modelling, theoretical modelling and regression for F3.1–F3.9 re-
⎪15.9457 + 2.3136 tan(0.5139𝑘𝑙 − 0.7384), spectively. They show excellent agreement for all these results between
⎪ the theoretical and regression solutions.
⎪ if 𝑘𝑙 ∈ [4.18, 𝜋∕0.7).
⎩
4. Intelligent evaluation method
3.3.2. Mode 2 buckling
From Fig. 4, for Mode 2 buckling, the effective length of half of The principal purpose of the current section is to develop the
the main member in the stayed beam–column is between that of the intelligent method for evaluating the nonlinear performance of the pre-
both ends pined and one end pinned and one end fixed cases. Hence, stressed stayed beam–columns using machine learning, such as nonlin-
𝑛 = [0.7, 1], 𝑘𝑙 = [𝜋, 𝜋∕0.7] ≈ [3.142, 4.488], rearranging Eq. (33) gives: ear failure mode and ultimate load. The preceding regression solutions
⎛ ⎞ of the linear buckling load are adopted for the current work.
𝐴s 𝑙2 ⎜ cos 𝛽u sin2 𝛽u cos 𝛽l sin2 𝛽l ⎟
(43)
𝐼c ⎜⎜ 1 + 2𝐴s 𝑙2 sin3 𝛽 ⎟.
𝐷2 = +
2𝐴 𝑙2 4.1. Data and features
u 1 + 3𝐼s sin3 𝛽l ⎟
⎝ 3𝐼a a ⎠
Substituting Eq. (38) to Eq. (43) gives: To obtain the data to train and test machine learning models, FE
analysis for the aforementioned 120 cases defined in Section 2.1 is
⎛ 2
cos 𝛽l sin2 𝛽l ⎞⎟ conducted. In terms of the prestressing level, from [33,35], 𝑇u,i sin 𝛽u =
𝛺 ⎜ cos 𝛽u sin 𝛽u
𝐷2 = + . (44)
𝑛2 𝜀s,y ⎜ 1 + 2𝛺 sin3 𝛽u 1 + 2𝛺 sin3 𝛽l ⎟ 𝑇l,i sin 𝛽l is applied for zero pre-cambering, and 7 prestressing levels
⎝ 𝑛 𝜀s,y 𝑛 𝜀s,y ⎠ 𝑇l,i ∕𝑇opt1 = {0, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0} are adopted. Hence, 120 × 7 =
Similar to Mode 1 buckling, 𝐷2 = [1.45, 23.29] can be determined 840 models are studied. To determine the nonlinear failure mode
numerically, and 𝐷2 = [1.45, 25.00] corresponding to 𝑘𝑙 = [3.50, 4.33] and ultimate load considering interactive buckling using FE mod-
can be adopted as the required range of 𝐹2,theory to determine 𝐹2,regression . elling, from [34,35], 3 imperfection combinations listed in Table 3
Applying NLS, the following expression is obtained: are adopted. From the existing work on the nonlinear stability of
the prestressed stayed beam–columns [33–35], the features strongly
𝐹2,regression = − 2.3799 + 3.2130 tan(0.6999𝑘𝑙 − 1.5753),
(45) linked to the nonlinear failure mode and ultimate load of the stayed
if 𝑘𝑙 = [3.50, 4.33). beam–columns are described as follows:
The comparison between the theoretical and regression functions
Structural parameters – the relative stiffness between the cable and
is shown in Fig. 9. It can be seen that both the theory and regression
main member 𝑟s = 𝜙s ∕(2𝑟c ), crossarm length ratio 𝛼l and 𝛼.
functions show excellent agreement. The absolute error for 𝐹1,regression
is mostly less than 3%; the absolute error for 𝐹2,regression is mostly less Prestressing levels – the relative prestressing force 𝑟T = 𝑇l,i ∕𝑇opt1 .
than 0.4%. This level of accuracy is sufficient for engineering practice.
Interactive buckling indicator – the difference ratio of mode buck-
3.4. Buckling loads 𝐶
ling loads 𝑟P = (𝑃1,regression 𝐶
− 𝑃2,regression 𝐶
)∕𝑃2,regression .

Replacing 𝐹1,theory and 𝐹2,theory in Eqs. (17) and (32) with 𝐹1,regression It should be noted that, all the aforementioned features and predicted
and 𝐹2,regression in Eqs. (42) and (45) give: variables are normalized to the ratios based on the explicit variables,
which would make the proposed intelligent method insensitive to the
𝐷𝑖 = 𝐹𝑖,regression = 𝑏𝑖,1 + 𝑏𝑖,2 tan(𝑏𝑖,3 𝑘𝑙 + 𝑏𝑖,4 ), (46) scale of the structures.

8
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 10. Buckling load comparison from FE modelling, theoretical modelling and regression expressions for cases F3.1–F3.9: 𝑃 C is the buckling load for Mode 1 or 2; ‘FE’ represents
the results from the linear buckling analysis within ABAQUS; ‘Theoretical’ represents the results from the theoretical implicit equations; ‘Regression’ represents the results from
the currently obtained explicit regression equations.

Fig. 11. Working flow for training and testing the model of nonlinear failure mode evaluation using XGBoost. 5 features belonging to 3 categories are input, and the failure mode
among DM1, UM1 and DI is output from this XGBoost classifier.

Fig. 12. Working flow for training and testing the model of ultimate load evaluation using ANN: 𝑁 is the number of neurons in each layer; 𝛾 represents 𝛾u and 𝛾d in Eq. (50),
and 𝑃 C represents 𝑃𝑖,regression
𝐶
in Eq. (50).

9
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 13. Comparison of the ultimate load between from traditional and proposed ANN methods. The dashed line is 𝑃u,predict = 𝑃u,actual , and the point nearer to this line indicates
more accurate prediction; the point under this line indicates safe prediction.

4.2. Machine learning models is DM1; if the buckling direction depends on the direction of 𝜇1 and the
ultimate load from the imperfection combination 2 is lower than that
In the current work, XGBoost and ANN are adopted. The algorithm from the imperfection combination 1, the nonlinear failure mode is DI.
of XGBoost is from Ensemble Learning, and it is a powerful machine Based on the FE results, the nonlinear failure mode of each model
learning model to conduct classification. Ensemble method helps im- can be determined from the preceding definitions. The nonlinear failure
prove prediction accuracy by combining predictions from multiple mode is used as the target value to train the machine learning model.
models, which leads to a better and robust predictive performance.
Since there are 3 failure modes, i.e. UM1, DM1 and DI, 3-classification
ANN is a powerful and popular deep learning model. It can learn from
XGBoost model is adopted. In terms of the data set, 30% of the 840
high-dimensional data, and sometimes complicated feature engineering
modelling results are used as the test set, and 70% of the 840 modelling
is not necessary. ANN can solve complex nonlinear problems by increas-
results are used as the training set. To optimize the model parameters,
ing model depth. Currently, Python is used to develop program, and the
grid-search method is used, and 4-fold cross-validation is applied in
open-source Python libraries XGBoost and TensorFlow are adopted to
conduct XGBoost and ANN modelling respectively. the training set. The parameters corresponding to the highest mean
of the accuracy across the 4 folds would be adopted as the best
4.3. Nonlinear failure mode evaluation parameters. The working flow for training and testing this XGBoost
model is illustrated in Fig. 11. In the test set, the overall prediction
In [34], 3 nonlinear failure modes for mono-symmetric stayed accuracy is 96.32%, and the confusion matrix is shown in Table 4. It
beam–columns were found from nonlinear analytical modelling, i.e. can be seen that the percentages in the diagonal positions are higher
Upward Mode 1 (UM1), Downward Mode 1 (DM1) and Downward than 90%, and the accuracy of the prediction for interactive buckling
Interactive (DI). For each of the 840 models, 3 imperfection combi- is 98.2%.
nations listed in Table 3 are applied in the FE analysis. If the upward
buckling occurs regardless of the imperfection shape and direction, the
4.4. Ultimate load evaluation
nonlinear failure mode is UM1; if the buckling direction depends on the
direction of the mode 1 imperfection (i.e. positive or negative 𝜇1 ) and
the ultimate load from the imperfection combination 1 is lower than Currently, the ultimate load evaluation considers upward and down-
that from the imperfection combination 2, the nonlinear failure mode ward buckling separately [34]. The ultimate loads of upward buckling

10
K. Wu et al. Engineering Structures 255 (2022) 113902

Fig. 14. Comparison of the prediction distribution between from traditional and proposed ANN methods. ‘pdf’ and ‘cdf’ are ‘Probability Density Function’ and ‘Cumulative
Distribution Function’ respectively; the dashed line is 𝑃u,predict ∕𝑃u,actual = 1.2. Higher levels of pdf near to 1.0 indicate better model, and lower levels of pdf at the right-hand side
of the dashed line indicate safer model with higher reliability.

Table 4 to the nonlinear failure mode evaluation, 70% and 30% of the 840
Confusion matrix of nonlinear failure mode evaluation in test set. ‘A’ represents the
modelling results are used as the train and test sets respectively.
actual failure mode from FE analysis, and ‘P’ represents the predicted failure mode
from XGBoost model. The optimization criteria for the model parameters is self-defined,
Accuracy P = DM1 P = DI P = UM1 including prediction accuracy (𝐴𝑐𝑐) and coefficient of variation (𝐶𝑜𝑉 )
thus:
A = DM1 95.10% 2.94% 1.96%
A = DI 0.90% 98.20% 0.90% ∑ 𝑃u,predict
|𝑃 − 1| 𝜎𝑃 ∕𝐴
A = UM1 5.13% 2.56% 92.31% u,actual
𝐴𝑐𝑐 = 1 − , 𝐶𝑜𝑉 = , (51)
𝑛 𝜇𝑃 ∕𝐴
where 𝑛 is the size of the test set; 𝑃u,predict and 𝑃u,actual are the ultimate
𝑃u,u and the ultimate load of downward buckling 𝑃u,d are expressed loads from machine learning and FE modelling respectively; 𝜇𝑃 ∕𝐴 and
respectively as: 𝜎𝑃 ∕𝐴 are the mean and volatility of 𝑃u,predict ∕𝑃u,actual in the test set re-
spectively. In addition, two ratios, i.e. overestimated ratio 𝑅𝑃 ∕𝐴>1 (the
𝐶 𝐶 𝐶
𝑃u,u = 𝛾u × 𝑃1,regression , 𝑃u,d = 𝛾d × min{𝑃1,regression , 𝑃2,regression }, (50) ratio of overestimation in the test set) and excessively overestimated
ratio 𝑅𝑃 ∕𝐴>1.2 (the ratio of overestimation higher than 20% in the
where 𝛾u and 𝛾d are the reduction factors to the buckling loads for
test set), are also defined. 𝐴𝑐𝑐, 𝐶𝑜𝑉 , 𝜇𝑃 ∕𝐴 , 𝑅𝑃 ∕𝐴>1 and 𝑅𝑃 ∕𝐴>1.2 are
upward and downward buckling respectively. Since there is an upward
adopted as the indexes to optimize the model parameters and evaluate
force within the crossarm, the failure mode of all the upward buck-
𝐶 the proposed ultimate load prediction modelling.
ling is UM1, and 𝑃1,regression is used as the basis of the ultimate load
For the ANN modelling, 30% data of the training set are used as
evaluation; however, for downward buckling, the failure mode may the validation set. Apart from the input and output layers, there are 3
𝐶
be DM1 or DI, and min{𝑃1,regression 𝐶
, 𝑃2,regression } is used as the basis to hidden layers and the numbers of neurons in each layer are 64, 32, 16
take interactive mode into account in the ultimate load evaluation. respectively. Mean Square Error is adopted as the loss function, and the
Based on the FE results, the ultimate loads of each model with different activation function is ‘Relu’. Two decimal places from 𝛾u and 𝛾d rounded
imperfection combinations listed in Table 3 are obtained. The ultimate down are adopted as the target values. The working flow for training
load of each model with imperfection combination 3 is used as 𝑃u,u , and and testing this ANN model is illustrated in Fig. 12.
the minimum ultimate load of each model between with imperfection The comparison of the ultimate load prediction between from the
combination 1 and with imperfection combination 2 is used as 𝑃u,d . 𝛾u traditional and proposed ANN methods is shown in Fig. 13. From
and 𝛾d of each model can be determined from the preceding definitions, Fig. 13, it can be seen that the prediction of the ultimate load from the
and they are used as the target values to train the machine learning traditional design methods in [20,21] is less accurate than that from
model. ANN model is applied to the ultimate load prediction. Similar ANN, since in Fig. 13(a) and (b) the number of points located near to

11
K. Wu et al. Engineering Structures 255 (2022) 113902

Table 5 Acknowledgements
Metrics comparison for the ultimate load prediction.
𝐴𝑐𝑐 𝐶𝑜𝑉 𝜇𝑃 ∕𝐴 𝑅𝑃 ∕𝐴>1.2 𝑅𝑃 ∕𝐴>1 The authors gratefully acknowledge the financial supports provided
Wadee [20] 0.755 0.255 1.100 0.361 0.578 by National Key Research and Development Program of China [Grant
Tankova [21] 0.825 0.225 1.049 0.239 0.409
No. 2020YFD1100400 and 2017YFB0304701], Natural Science Foun-
ANN for upward 0.941 0.065 0.948 0.000 0.177
ANN for downward 0.943 0.051 0.949 0.005 0.106
dation of China [Grant No. 52142804, 51408150], National Ministry
of Finance Program, China [Grant No. (2013) 235] and Fundamental
Research Funds for the Central Universities, China.

the line of 𝑃u,predict = 𝑃u,actual is smaller than that in Fig. 13(c) and (d). Appendix A. Supplementary data
In particular, in Fig. 13(a) and (b) the number of points located in the
safe area is smaller than that in Fig. 13(c) and (d), which indicates Supplementary material related to this article can be found online
that the proposed ANN method is safer than the traditional methods. at https://doi.org/10.1016/j.engstruct.2022.113902.
The comparison of the prediction distribution between from traditional
and proposed ANN methods is shown in Fig. 14. Fig. 14 validates the References
conclusions from Fig. 13 further. It can be seen that 𝑅𝑃 ∕𝐴>1.2 is high in
Fig. 13(a) and (b), but 𝑅𝑃 ∕𝐴>1.2 approximates to zero in Fig. 13(c) and [1] Chu KH, Berge SS. Analysis and design of struts with tension ties. ASCE J Struct
(d). This phenomenon indicates that the ANN model has high reliability Div 1963;89:127–63.
in ultimate load prediction. The metrics comparison for the ultimate [2] Smith RJ, Ellis JS, McCaffrey GT. Buckling of a single-crossarm stayed column.
load prediction is listed in Table 5, which indicates that the ANN ASCE J Struct Div 1975;101(1):249–68.
[3] Hafez HH, Temple MC, Ellis JS. Pretensioning of single-crossarm stayed columns.
model perform better than the traditional method. It should be noted ASCE J Struct Div 1979;105(2):359–75.
that in the intelligent method for the ultimate load evaluation, the [4] Temple MC. Buckling of stayed columns. ASCE J Struct Div 1977;103(4):839–51.
adopted input features and predicted variables are from the mechanical [5] Wong KC, Temple MC. Stayed column with initial imperfection. ASCE J Struct
studies on the behaviour of the prestressed stayed beam–columns, Div 1982;108:1623–40.
and they can represent the mechanical relationship between the input [6] Osofero AI, Wadee MA, Gardner L. Experimental study of critical and post-
buckling behaviour of prestressed steel stayed columns. J Construct Steel Res
and output. Similar to other machine learning methods, the proposed
2012;79:226–41.
intelligent method has no explicit expression but high accuracy, while [7] Osofero AI, Wadee MA, Gardner L. Numerical studies on the buckling resistance
the traditional design method has explicit expression which is much of prestressed stayed columns. Adv Struct Eng 2013;16:487–98.
easier to be understood by engineers but low accuracy. [8] Serra M, Shahbazian A, Simões da Silva L, Marques L, Rebelo C, Vellasco PCGdS.
A full scale experimental study of prestressed stayed columns. Eng Struct
5. Concluding remarks 2015;100:490–510.
[9] Lv G, Li P, Zhang C, Sun X, Chen E. Experimental investigation into the
buckling behaviour of cable-stiffened steel columns. J Construct Steel Res
An analytical model, which allows the direct evaluation for the 2021;183:106753.
buckling loads of both doubly-symmetric and mono-symmetric stayed [10] Li P, Wadee MA, Yu J, Christie NG, Wu M. Stability of prestressed stayed steel
beam–columns under purely axial compression, was developed. The columns with a three branch crossarm. J Construct Steel Res 2016;122:274–91.
implicit and explicit solutions of buckling loads were obtained. The so- [11] Zhou P, Guo YL, Pi Y. Numerical investigations of multiple cross-arm pre-
tensioned cable stayed BRBs with pin-ended stays. J Construct Steel Res
lutions showed excellent agreement with the buckling eigenvalues from
2018;148:206–32.
the linear buckling analysis in ABAQUS. Subsequently, XGBoost and [12] Lapira L, Wadee MA, Gardner L. Stability of multiple-crossarm prestressed stayed
ANN were applied to develop the models for evaluating the nonlinear columns with additional stay systems. Structures 2017;12:227–41.
failure mode and ultimate load of prestressed stayed beam–columns [13] Chan SL, Shu GP, Lü ZT. Stability analysis and parametric study of pre-stressed
respectively. XGBoost model can evaluate the nonlinear failure mode stayed columns. Eng Struct 2002;24:115–24.
[14] Saito D, Wadee MA. Post-buckling behaviour of prestressed steel stayed columns.
including interactive buckling accurately, and the overall accuracy of
Eng Struct 2008;30(5):1224–39.
prediction is 96.32%. For the ultimate load evaluation, the ultimate [15] Saito D, Wadee MA. Buckling behaviour of prestressed steel stayed columns with
load prediction using ANN is higher than 94%, which can consider imperfections and stress limitation. Eng Struct 2009;31(1):1–15.
interactive buckling effect. The analytical solutions of the buckling [16] Saito D, Wadee MA. Numerical studies of interactive buckling in prestressed steel
loads can be applied to evaluate the critical buckling between the stayed columns. Eng Struct 2009;31(2):432–43.
symmetric and anti-symmetric modes independent to prior FE analysis. [17] Wadee MA, Gardner L, Hunt TA. Buckling mode interaction in prestressed stayed
columns. Proc Inst Civ Eng – Struct Build 2013;166:403–12.
The machine learning method for evaluating the ultimate load can be
[18] Yu J, Wadee MA. Mode interaction in triple-bay prestressed stayed columns. Int
used to develop the intelligent design method for stayed beam–columns J Non-Linear Mech 2017;88:47–66.
in future. [19] Li P, Liu X, Zhang C. Interactive buckling of cable-stiffened steel columns with
pin-connected crossarms. J Construct Steel Res 2018;146:97–108.
CRediT authorship contribution statement [20] Wadee MA, Gardner L, Osofero AI. Design of prestressed stayed columns. J
Construct Steel Res 2013;80:287–98.
[21] Tankova T, da Silva LS, Martins JP. Stability design of cable-stayed columns:
Kaidong Wu: Conceptualization, Methodology, Software, Formal
Existing methods and future perspectives. Steel Constr 2019;12(4):309–17.
analysis, Investigation, Data curation, Writing – original draft, Visu- [22] Wang H, Li P, Wu M. Crossarm length optimization and post-buckling analysis
alization, Writing – review & editing. Xuhong Qiang: Conceptualiza- of prestressed stayed steel columns. Thin-Walled Struct 2019;144:106371.
tion, Methodology, Investigation, Data Curation, Project administra- [23] Li P, Wang H. A novel strategy for the crossarm length optimization of
tion, Funding acquisition, Writing – review & editing. Zhe Xing: Con- psscs based on multi-dimensional global optimization algorithms. Eng Struct
2021;238:112238.
ceptualization, Methodology, Software, Formal analysis, Investigation,
[24] Pichal R, Machacek J. Buckling and post-buckling of prestressed stainless steel
Data curation, Writing – review & editing. Xu Jiang: Conceptualiza- stayed columns. Eng Struct Technol 2017;9(2):63–9.
tion, Methodology, Investigation, Funding acquisition, Writing – review [25] Wang H, Li P, Jian B, Meng L. Investigation of buckling resistance of
& editing. cable-stiffened aluminium alloy columns. Thin-Walled Struct 2021;165:107946.
[26] Gosaye J, Gardner L, Wadee MA, Ellen ME. Tensile performance of prestressed
Declaration of competing interest steel elements. Eng Struct 2014;79:234–43.
[27] Gosaye J, Gardner L, Wadee MA, Ellen ME. Compressive behaviour and design
of prestressed steel elements. Structures 2016;5:76–87.
The authors declare that they have no known competing finan- [28] Li P, Yang Y, Yuan J, Jia B. Numerical investigation into prestressed stayed
cial interests or personal relationships that could have appeared to steel box section columns under eccentric loading. J Construct Steel Res
influence the work reported in this paper. 2019;159:1–12.

12
K. Wu et al. Engineering Structures 255 (2022) 113902

[29] Hyman P, Osofero AI. Behaviour of eccentrically loaded prestressed stayed [39] Sun H, Burton HV, Huang H. Machine learning applications for building
columns with circular hollow sections. Adv Struct Eng 2020;23(13):2813–21. structural design and performance assessment: state-of-the-art review. J Build
[30] Feng P, Hu L. Steel columns strengthened/reinforced by prestressed cfrp Eng 2020;101816.
strips: concepts and behaviors under axial compressive loads. Compos Struct [40] Zhu S, Ohsaki M, Guo X, Zeng Q. Shape optimization for non-linear buckling
2019;217:150–64. load of aluminum alloy reticulated shells with gusset joints. Thin-Walled Struct
[31] Hu L, Feng P. Prestressed cfrp-reinforced steel columns under axial and eccentric 2020;154:106830.
compression. Compos Struct 2021;268:113940. [41] Zhu S, Ohsaki M, Guo X. Prediction of non-linear buckling load of imperfect
[32] Hu L, Feng P, Meng Y, Yang J. Buckling behavior analysis of prestressed reticulated shell using modified consistent imperfection and machine learning.
cfrp-reinforced steel columns via fem and ann. Eng Struct 2021;245:112853. Eng Struct 2021;226:111374.
[33] Wu K, Wadee MA, Gardner L. Stability and ultimate behaviour of prestressed [42] Xu Y, Zheng B, Zhang M. Capacity prediction of cold-formed stainless steel
stayed beam–columns. Eng Struct 2019;201:109723. tubular columns using machine learning methods. J Construct Steel Res
[34] Wu K, Wadee MA, Gardner L. Interactive buckling in prestressed stayed 2021;182:106682.
beam–columns. Int J Mech Sci 2020;174:105479. [43] Saito D, Wadee MA. Optimal prestressing and configuration of stayed columns.
[35] Wu K, Wadee MA, Gardner L. Prestressed stayed beam–columns: Sensi- Proc Inst Civ Eng – Struct Build 2010;163(5):343–55.
tivity to prestressing levels, pre-cambering and imperfections. Eng Struct [44] Yu J, Wadee MA. Optimal prestressing of triple-bay prestressed stayed columns.
2021;226:111344. Structures 2017;12:132–44.
[36] Wu K, Xing Z. Stability of imperfect prestressed stayed beam–columns under [45] Bethea RM. Statistical methods for engineers and scientists, Vol. 144. CRC Press;
combined axial load and bending. Eng Struct 2021;245:112891. 1995.
[37] Krishnan S. Cable-stayed columns and their applications in building structures. [46] BS EN 1993-1-1. Eurocode 3: Design of steel structures - Part 1-1: General rules
J Build Eng 2020;27:100984. and rules for buildings. British Standards Institution; 2005.
[38] ABAQUS, version 6.14 documentaion. Providence, RI, USA: Dassault Systèmes
Simulia Corp.; 2016.

13