You are on page 1of 16

Advanced Engineering Informatics 46 (2020) 101154

Contents lists available at ScienceDirect

Advanced Engineering Informatics


journal homepage: www.elsevier.com/locate/aei

An optimized combination prediction model for concrete dam deformation T


considering quantitative evaluation and hysteresis correction
Qiubing Rena, Mingchao Lia, , Lingguang Songb, Han Liuc

a
State Key Laboratory of Hydraulic Engineering Simulation and Safety, Tianjin University, Tianjin 300350, China
b
Department of Construction Management, University of Houston, Houston, TX 77204, USA
c
Bei Fang Investigation, Design & Research CO. LTD, Tianjin 300222, China

ARTICLE INFO ABSTRACT

Keywords: Certain degree of deformation is natural while dam operates and evolves. Due to the impact of internal and
Dam deformation prediction external environment, dam deformation is highly nonlinear by nature. For dam safety, it is of great significance
Combination modeling to analyze timely deformation monitoring data and be able to predict reliably deformation. A comprehensive
Hyper-parameter optimization review of existing deformation prediction models reveals two issues that deserves further attention: (1) each
Synthetic evaluation
environmental influencing factor contributes differently to deformation, and (2) deformation lags behind en­
Hysteresis correction
vironmental factors (e.g., water level and air temperature). In response, this study presents a combination de­
formation prediction model considering both quantitative evaluation of influencing factors and hysteresis cor­
rection in order to further improve estimation accuracy. In this study, the complex relationship in deformation
prediction is effectively captured through support vector machine (SVM) modeling. Furthermore, a modified
fruit fly optimization algorithm (MFOA) is presented for SVM hyper-parameter optimization. Also, a synthetic
evaluation method and a hysteresis quantification algorithm are introduced to further enhance the MFOA-SVM-
based model in regards to contribution quantification and phase correction respectively. The accuracy and
validity of the proposed model is evaluated in a concrete dam case, where its performance is compared with
other existing models. The simulated results indicated that the proposed nonlinear MFOA-SVM model con­
sidering both quantitative evaluation and hysteresis correction, abbreviated as SEV-MFOA-SVM, is more accu­
rate and robust than conventional models. This novel model also provides an alternative method for predicting
and analyzing dam deformation and evolution behavior of other similar hydraulic structures.

1. Introduction operational state, and it is relatively easy to observe and measure [4].
Deformation prediction modeling is a fundamental for modern dam
Dams play an important role in the socio-economic growth in many safety evaluation. Actual deformation can be compared with the model-
countries because of their enormous benefits in water supply, flood predicted value as the baseline to detect deviation, if any, of the dam
control, power generation, irrigation and navigation. Most of the dams from its expected performance. In the case of a large deviation beyond a
built around the world with a height above 100 m are concrete dams. threshold value, discovered risk alarm is confirmed, and remedial
Statistically, there are 16 concrete dams among the 30 tallest dams in measures can be taken timely to prevent failures [5]. Therefore, a re­
China ranked by dam height [1,2]. Despite of its tremendous benefits, liable and accurate safety monitoring and predication model is critical
the failure of a dam would be catastrophic to human lives and property, for concrete dam.
national economy and ecological environment in reservoir regions and Dam deformation is influenced by many environmental factors and
beyond. To closely monitor dam structural health, especially the de­ material mechanical properties, including hydrostatic pressure, tem­
velopment of deformation, seepage and crack in response to environ­ perature and time effects due to concrete aging, bedrock creep, etc. [6].
mental factors (e.g., water level and air temperature), many monitoring A safety monitoring model is usually formulated based on the causal
devices and systems are commonly installed inside and around the dam relationship between these influencing factors and corresponding dam
[3]. Among various types of such dam structural responses, horizontal deformation. Table 1 shows three common deformation prediction
deformation is considered as a reliable and intuitive indicator of dam modeling methods. The traditional deterministic modeling method uses


Corresponding author.
E-mail address: lmc@tju.edu.cn (M. Li).

https://doi.org/10.1016/j.aei.2020.101154
Received 1 October 2019; Received in revised form 7 June 2020; Accepted 3 August 2020
1474-0346/ © 2020 Elsevier Ltd. All rights reserved.
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

Table 1 algorithm [27], particle swarm optimization (PSO) [17], salp swarm
Three common deformation prediction modeling methods. algorithm [3], orthogonal design [28,29], etc. Nonetheless, the above
Model Common modeling methods optimization algorithms are not only complex to implement, but also
difficult to understand. Inspired by these facts, there are also some
Deterministic model FEM and hybrid modeling simple and efficient optimization techniques that can be utilized for
Statistical model HST and HTT
dam behavior modeling. Recently, a novel intelligent optimization al­
Artificial intelligence model ANN, SVM, ELM, ANFIS, RF, RBFN, BRT and GP
gorithm based on the foraging behavior of drosophila, named fruit fly
optimization algorithm (FOA), was proposed by Pan [30]. Unlike
numerical simulation to describe and predict dam deformation ac­ others, FOA has certain outstanding merits, such as shorter program
cording to the structural form and mechanical parameters of a dam and code, stronger searching ability, and higher solution precision. FOA has
its foundation [7]. This model-based approach is often used during been successfully applied to various fields such as power load fore­
design, construction, and initial storage stages of a new dam with the casting [31], service satisfaction analysis [32], etc. These FOA-related
absence of any significant long-term monitoring data. However, its works demonstrated that FOA is effective for hyper-parameter optimi­
practical implementation can be cumbersome and time-consuming, and zation of MLAs. As such, FOA is introduced here to optimize SVM
further, assumptions and simplifications with regards to geometric and hyper-parameters, and ultimately, a combination deformation predic­
boundary conditions negatively impact the prediction accuracy [8]. tion model based on modified FOA and SVM is proposed in our work.
Another commonly used data-driven modeling method is statistical It should be noted that the regression model such as SVM alone has
models, which usually involve hydrostatic-season-time (HST) modeling limited capability in modeling the complex potential relationship be­
and hydrostatic-temperature-time (HTT) modeling [9,10]. The multi­ tween dam deformation and its influencing factors for several reasons.
variate linear regression, stepwise regression (SR), and partial least The magnitude of each influencing factor and effect on dam deforma­
squares regression are extensively adopted to solve unknown model tion is dynamically changing during service [33]. The leading factors of
coefficients. The merits of these statistical models are explicit physical deformation at various monitoring points even in the same dam block
interpretation, simple model structure, and fast execution. Meanwhile, are also different. For example, the monitoring point of the dam crest is
most statistical models are linear regression in nature, which limits more sensitive to air temperature than that of the dam body. The de­
their fitting accuracy and reliability. Some hybrid models [11] are also formation prediction model should be able to treat various factors (i.e.,
applied in engineering practice, which utilize finite element method input variables) differently in order to realistically model real-world
(FEM) to simulate the hydrostatic component, while the temperature systems. However, the input weights assigned to different normalized
and the aging component are still calculated by statistical models. This factors by SVM are almost identical [34]. To this end, Dai et al. [33]
substantiates the effectiveness of the integrated approach on improving adopted linear regression coefficients from SR to assign weights to each
prediction accuracy. factor, and set up a statistical model optimized RF regression model for
Concrete dam is a dynamic and complex system, including dam dam deformation monitoring. Li et al. [34] introduced mutual in­
body, foundation, and the surrounding environment. Due to the special formation to reflect the nonlinear contribution of different factors to
structural morphology and the complex external environment, the de­ deformation. These methods only focus on the relevance between en­
formation behavior, as a comprehensive reflection of dam integrity, is vironmental factors and dam deformation without considering in­
characterized by nonlinearity, uncertainty, and time-variation [12]. dividual factor behavior and the correlation among different factors. An
Furthermore, the exact relationship between dam deformation and its evaluation method that synthetically considers not only the variables
influencing factors is unclear. Hence, it is challenging to predict dam themselves but also the relationship among variables is thus desired. To
deformation accurately. With the rapid development of artificial in­ the best of our knowledge, this problem is rarely considered in the past
telligence, data-driven machine learning algorithms (MLAs) gradually studies of dam deformation prediction model using SVM. Therefore,
replace linear regression methods in developing dam safety monitoring this study will be one of the first to explore the issue of synthetic eva­
models. These modeling approaches include artificial neural networks luation of different factors for dam deformation prediction.
(ANN) [13,14], support vector machine (SVM) [15–17], extreme The dam structural behavior is not synchronous with changing
learning machine (ELM) [18], adaptive network-based fuzzy inference water levels, air temperatures, and other environmental factors
system (ANFIS) [19], random forest (RF) [20], radial basis function [35–37], i.e. dam deformation lags behind. The dam deformation at any
networks (RBFN) [21], boosted regression trees (BRT) [22] and Gaus­ given moment is related not only to the current water level (or tem­
sian process (GP) regression [23], as summarized in Table 1. These perature), but also to the previous water level (or temperature).
advanced nonlinear modeling methods can simulate the complex in­ Therefore, the hysteresis effects of dam deformation to water level and
teraction between variables, and resist the noise interference in mon­ air temperature must be taken into account in the monitoring model. As
itoring data without presupposition, which is particularly suitable for the hysteresis mechanism is too complex to model explicitly, it is
the interpretation of dam behavior [23]. In many instances, MLAs can usually simplified in actual practice. The statistical models typically
provide more accurate predictions than statistical models. Literature take hysteresis into consideration by adding a general lag term [36,37],
review shows that ANN and SVM are the two most widely used MLAs in while deterministic models usually consider this hysteresis effect by
dam safety monitoring [21]. Compared with ANN, SVM has a clear delaying the relevant loads for a randomly determined duration. In
advantage in solving small-sample, high-dimensional, and nonlinear contrast, the artificial intelligence models empirically introduce the
problems due to its solid theoretical basis [3,23]. In real-world appli­ average water level (or temperature) of the previous 30, 60 or 90 days
cations, SVM has also achieved excellent results, such as landslide as additional input variables [21]. This study presents an alternative
prevention [24], patch detection [25], and pattern recognition [26]. algorithm to quantify hysteresis. The deformation time series is visually
The prediction accuracy of SVM depends on the reasonable setting of staggered from water level and air temperature series, and it is rea­
hyper-parameters, but there is no consensus on the appropriate reg­ sonable to infer that the preceding most similar and closest environ­
ularization constant to avoid overfitting, and on selecting the optimal mental factors are regarded as representative hysteresis variables via
kernel function and associated parameters. The meta-heuristic optimi­ phase correction. The time span of corrected phase is collectively
zation algorithm may be employed to adaptively search for the optimal known as hysteresis hereinafter. According to our preliminary tests, the
hyper-parameters of SVM, thus constructing a combination prediction simplified method is useful for grasping the variation trend of de­
model with powerful performance. Different algorithms have been used formation with changing water levels and temperatures.
to tune SVM for mathematical modeling so far, which include genetic To sum up, a novel SVM-based combination prediction model for
concrete dam deformation is presented in this paper, including hyper-

2
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

parameter optimization, quantitative evaluation, and hysteresis cor­ 3. Methodology


rection. The corresponding modified FOA, synthetic evaluation method,
and hysteresis quantification algorithm are presented to improve the 3.1. Hysteresis quantification algorithm
nonlinear mining capability of SVM. The performance of the combi­
nation prediction model is validated in a case study. The simulated This paper presents a quantitative algorithm for calculating the
results show that the proposed model is a viable option for dam de­ hysteresis time of dam deformation due to influencing factors (up­
formation prediction. stream water level and air temperature) by studying fluctuation and
The remaining of this paper is organized as follows. The literature phase of actual monitoring data. Taking the dam deformation (δ) and
review of SVM-based prediction models for dam deformation is de­ water level (H) for example, both are time series data composed of
picted in Section 2. In Section 3, the implementation procedure of the trend term (Tt), periodic term (Pt), and irregular variation term (It). The
hysteresis quantification algorithm and synthetic evaluation method, basic idea of hysteresis quantification algorithm is to extract Tt from δ
and the background knowledge of SVM and modified FOA are briefly and H, conduct additive reconstruction of the remaining Pt and It, and
introduced. The construction process of the proposed combination then carry out sliding phase matching for the periodicity of the two
prediction model is also explained in Section 3. Section 4 describes the reconstructed series, thus quantifying the hysteresis time of δ due to H
case study application and simulation scheme in details. Section 5 by measuring the matching degree. The specific steps of this algorithm
presents the simulation and analysis results of the proposed model and are as follows:
the comparison with other models. Finally, conclusions and future work Step 1: Normalization and interpolation. δ and H series are nor­
are summarized in Section 6. malized into [0, 1] to reduce the effect of order of magnitude difference,
expressed as Eq. (1). The piecewise cubic Hermite spline interpolation
[46] has first-order continuity, and its interpolation curve is relatively
2. Literature review on SVM-based prediction models for dam smooth and conformal due to the full use of actual monitoring data.
deformation This interpolation is used to increase the data density of δ and H series,
thus reducing the time granularity of hysteresis quantification.
The SVM method is widely used in dam behavior modeling and X Xmin
prediction. Many scholars at home and abroad have applied SVM to X norm =
Xmax Xmin (1)
dam deformation prediction and achieved satisfactory results. For ex­
ample, Song and Li [38] compared the performance of SVM and its where X and are the monitoring data before and after normal­
X norm
extension method least squares SVM (LSSVM) in dam deformation ization, and Xmax and Xmin are the maximum and minimum values in a
prediction. The outcomes showed that both algorithms have high pre­ data series, respectively.
diction accuracy, and LSSVM has better performance. Tabari and Sa­ Step 2: Decomposition and reconstruction. The moving average
nayei [39] directly implemented the SVM and ANN models for pre­ (MA) method, whose principle is shown in literature [47], is often used
diction of the intermediate block displacement of the dam crest. It was to separate Tt in the original series. MA can filter out high frequency
discovered that the prediction error in the SVM model is about 32% interference in δ and H, and retain useful low frequency trend in­
lower than the ANN model. Zhang et al. [27] and Wang et al. [40] formation. However, MA directly uses the measured data of the pre­
decomposed the dam deformation series into several components of ceding period to estimate the value at the next time point, which will
different scales through empirical mode decomposition and wavelet expand the hysteresis after trial calculation. For this purpose, a cor­
transform respectively, and used SVM-based models to predict each rection strategy that uses both the preceding and succeeding period
component. The prediction results of each component are superimposed data for estimation is proposed, see Eq. (2). When Tt is extracted from δ
to obtain the final deformation prediction results. Ranković et al. [15] and H by the corrected MA, the remaining two terms (Pt and It) are
and Kang et al. [3] modified the input variables of SVM from the de­ added to reconstruct new series named and H .
formation series itself and temperature effect respectively. It was found
X norm norm
n 1 +X n 3 + + Xtnorm + + X norm
that the modified approach can result in significant reduction in de­ t t n 1
t+
2 2 2
, if n is odd
formation prediction errors. Su et al. [41,42] constructed two mon­ Ft = n
X norm norm
n 2 +X n 4 + + Xtnorm + + X norm
itoring models of safety status based on SVM during structural re­ t
2
t
2
n
t+
2
, if n is even (2)
inforcement of dangerous dams, one focusing on time-varying n

identification and the other on real-time control. Some researchers where Ft is the estimated value at the next time point, n is the number of
employed SVM to study the chaotic characteristics of dam deformation monitoring data contained in a single moving average period, and
series. The dam safety prediction model considering chaotic char­ Xtnorm norm norm norm are measured values before and after
n 1 , Xt n 2 , …, Xt + n 1 , Xt + n
acteristics in prototype observations were built by combining SVM, 2 2 2 2
the time point t .
chaos theory, and PSO [43]. Similarly, Su et al. [17] utilized SVM with
Step 3: Sliding matching and hysteresis calculation. One re­
Morlet wavelet basis function, phase space reconstruction, and im­
constructed series ( ) stays the same, while the other (H ) moves in a
proved PSO to build the prediction model for dam deformation. Wei
fixed direction with a step-size of one. In the process of moving, cosine
et al. [44] presented an optimized prediction model for concrete dam
similarity method [48] is introduced to measure the similarity of two
deformation based on signal residual amendment for the uncertain
series, where the closer the cosine (see Eq. (3)) is to one, the more
chaos effect implied in residual series. The dynamic behavior of dam
similar the two series are. The position with the largest cosine similarity
deformation can also be characterized by SVM and its extension
is the phase hysteresis, and then the hysteresis time of to H can be
methods. Su et al. [16] raised a performance improvement method of
easily obtained through visualization, which is equivalent to the hys­
SVM-based model monitoring dam safety, thus advancing the modeling
teresis of δ with respect to H.
efficiency and forecasting ability. Li and Ren [34,45] proposed a dy­
namic monitoring model for dam deformation considering factor cor­ n
(Xinorm ·Yinorm)
relation and sequence similarity, in which SVM is a learning machine i=1
that simulates the causal relationship between influencing factors and cos( ) = n n
dam deformation. The above facts indicate that SVM has excellent (Xinorm )2 · (Yinorm )2
i=1 i=1 (3)
nonlinear mining and regression estimation capabilities under small-
sample and high-dimensional conditions, so that it is suitable for dam where cos( ) is the cosine of the angle between the two vectors, and
deformation prediction and analysis. Xinorm and Yinorm are the components of both vectors. Similarly, the

3
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

hysteresis time of dam deformation (δ) to air temperature (T) can also 3.3. MFOA-optimized SVM for dam deformation prediction
be calculated in a similar way.
3.3.1. Support vector machine (SVM)
SVM is a supervised learning algorithm based on statistical theory
3.2. Synthetic evaluation method
and structural risk minimization principle [52,53]. The basic idea is to
construct an optimal decision hyperplane to maximize the distance
A synthetic evaluation method is proposed to measure the in­
between the two different classes of samples closest to the hyperplane.
formation about the influencing factors themselves and the complex
For linear regression, the function f (x ) = x + b is used to estimate the
nonlinear relationship between dam deformation and these influencing
data (xi , yi ), x i n, y , i = 1, 2, , l . When the slack variables i
factors. This method comprehensively considers the fluctuation of each i
and i are introduced, if f (x ) can reasonably estimate all data (xi , yi ) in
factor, the correlation between different factors, and the relationship
the case of precision , then it can be transformed into a convex opti­
between dam deformation and various influencing factors in de­
mization problem:
termining the synthetic evaluation value (SEV) of each factor. Therein,
the degree of data fluctuation is the discreteness of data, which is 1
l
min 2 +C (i+ )
commonly expressed as the coefficient of variation (CV) [49] in sta­ ,b 2 i
i=1
tistics. The greater the CV is, the greater the discreteness of the data is,
and also more information the data contains. The correlation between yi ( x i + b) + i

different factors include linear and nonlinear relationships, e.g., over­ s. t. ( x i + b) yi + i


lapping information between two factors. The maximal information i, i 0, i = 1, 2, ,l (9)
coefficient (MIC) [50], which is adopted here to measure the degree of
information overlap, is a novel method based on information and where is the weight vector, C , C > 0 is the penalty factor, b is the
probability theory to detect nonlinear correlation between variables. offset, and is the insensitive loss function. Both C and ε are user-de­
The larger the MIC between two factors is, the higher degree of in­ termined hyper-parameters, and ε is generally set to 0.01 [54,55].
formation overlap is, indicating that the factors contain less in­ The Lagrangian function and dual variables are introduced to get
dependent information. The relevance between dam deformation and the function (L ).
various factors is also measured via MIC. It can be said that the syn­ l l
1
thetic evaluation method is the combination of CV and MIC. Assuming L= 2
2 +C (i+ i ) i( + i yi + xi + b)
that m sets of normalized data of different factors X norm
j , j = 1, 2, , m i=1 i=1

and a deformation series need to be comprehensively evaluated, the


l l
i ( + + yi xi b) ( +
computation process of SEV value is as follows: i=1
i
i=1
i i i i
(10)
Step 1: Calculating the coefficient of variation (CVj) of each influ­
encing factor. where i and i are Lagrangian multipliers, and i and i are the
parameters corresponding to i and i respectively.
CVj =
j
, j = 1, 2, ,m According to Karush-Kuhn-Tucker conditions, a quadratic pro­
X norm
j (4) gramming problem is acquired.

where j and X norm


j are the standard deviation and mean of the j-th 1
l l l
max ( )( j ) xi xj + ( ) yi ( + )
factor respectively. i, i
2
i, j = 1
i i j
i=1
i i
i=1
i i

Step 2: Calculating the correlation coefficients (CCij) between dif­ l


ferent factors using MIC. ( i i )=0
s. t . i=1
CCij = MIC (Xinorm , X norm
j ), i,j = 1, 2, ,m (5) i, i [0, C ] (11)
where and
Xinorm are the i-th and j-th factors respectively, and
X norm
j The linear regression function f (x ) can be obtained by solving Eq.
MIC (·) is the calculation formula of MIC, which can be found in lit­ (11).
erature [51]. l
Step 3: Calculating the independent information coefficient (ICj) of f (x ) = x + b = ( i i ) xi x + b
each factor by CCij value. i=1 (12)
m For nonlinear regression, the low-dimensional space is mapped nl
ICj = (1 CCij ), i,j = 1, 2, ,m to the high-dimensional space nh by kernel function
i, j = 1, i j (6) k (x i , x ) = (x i )· (x ) , and then the optimal fitting hyperplane can be
Step 4: Calculating the comprehensive information coefficient (ωj) found in nh . In order to avoid explicitly defining feature mappings,
of each factor by coupling CVj and ICj. kernel functions for SVM should satisfy Mercer's theorem. Among many
kernel functions, polynomial kernel and radial basis kernel function
CVj· ICj (RBF) completely satisfy this theorem. Compared with polynomial
j = m , j = 1, 2, ,m
kernel function, RBF has fewer parameters and more practical appli­
(CVj·ICj )
j=1 (7) cations, thus it is used in this research and the nonlinear regression
function with RBF can be described as:
Step 5: Calculating the relevance coefficient (RCj) between dam
l
deformation and each influencing factor, and derterming the synthetic f (x ) = · (x ) + b = ( ) k (x i , x ) + b
i i
evaluation value (SEVj). i=1
2 ),
k (xi , x ) = exp( x xi >0 (13)
SEVj = j + (1 ) RCj
, j = 1, 2, ,m
RCj = MIC (X jnorm , Y norm) where is the width parameter of the RBF.
(8)

where is the weight allocation coefficient that can be acquired by trial 3.3.2. Modified fruit fly optimization algorithm (MFOA)
calculation, and Y norm represents the corresponding dam deformation FOA is a new optimization algorithm that simulates the foraging
series. process of drosophila using their developed olfactory and visual organs

4
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

[56,57]. With the iteration of smell search and visual search, the FOA where i represents the population size, and
maxstep
population continues to evolve to determine the optimal solution to a k= , gen = 1, 2, , maxgen .
maxgen
problem. FOA has some clear advantages, such as simple implementa­ 1 + exp gen
2
tion and intuitive model structure with only a few parameters to tune. Step 3: Fitness function calculation. The fruit fly location (xi , yi ) is
However, FOA also has some disadvantages, such as the inability to substituted into the fitness function to find the smell concentration
solve an optimization problem in a domain containing negative values (smelli ) of the individual location. The fruit fly with the best smell
yet requiring a uniform global search of the solution space. In such concentration (note that smaller fitness function value represents better
cases, FOA can easily get stuck in the local optimal, thus limiting its individual in this paper), and the corresponding location among the
application in similar use scenarios. To address these issues, FOA can be fruit fly swarm are then determined.
modified with regards to the algorithm itself and its search strategy, as
described below. smelli = Function (x i , yi )
(1) FOA is limited due to the fact that candidate solutions can only [bestsmell bestindex ] = min(smelli ) (16)
be positive and gradually approach zero. This is because FOA in its
Step 4: Movement. The fruit fly swarm retains the best smell con­
original form specifies that the smell concentration judgement value is
centration value, and flies towards the best coordinate location (xb , yb )
the reciprocal of the distance. Thus in our research, we directly take the
using vision.
fruit fly location instead as the candidate solution of the smell con­
centration judgement value. This new mechanism for candidate solu­ smellbest = bestsmell
tion generation not only solves the inherent limitation of FOA, but also xb = x (bestindex )
guarantees the search uniformity. yb = y (bestindex ) (17)
(2) The ability of an optimization algorithm lies in the combination
of global search and local search. In the early stage of optimization Step 5: Iteration. Enter the iterative optimization to repeat the
process, the algorithm can have strong global search ability by relying implementation of Steps 2–4. The iteration stops when the smell con­
on large step-size, while in the later stage, the algorithm needs to en­ centration is no longer superior to the previous iterative results or the
hance the local search ability by reducing the step-size. This paper iterative number reaches a user-specified maximal iterative number.
presents a dynamic step-size search strategy based on Logistic function
and the novel step-size adjustment coefficient (k ), as illustrated in Eq. 3.3.3. Verification of modification effect
(15). The mapping relationship between step-size and iteration number Three commonly used benchmark functions of known optimal so­
is shown in Fig. 1. This method allows dynamic step-sizes, and im­ lutions (see Table 2), a unimodal function and two multimodal func­
proving the convergence accuracy. tions, are used to evaluate the optimization performance of MFOA [58].
With the above two improvement measures, the modified FOA's The performance of MFOA is also compared with that of FOA in its
(i.e., MFOA) implementation procedure can be summarized as follows. original form. Under the same conditions (sizepop = 20 and
Step 1: Parameter setting. The maximum iteration number maxgen = 100 ), MFOA and FOA are both applied to optimize the three
(maxgen ), the population size (sizepop ), the maximum search step-size benchmark functions, and the optimization results are shown in
(maxstep ), the initial location range of fruit fly swarm (LR), and the Table 3. Each function is optimized 10 times with MFOA and FOA se­
random flight distance range (FR) are set. Also, the initial fruit fly parately, and the mean and the standard deviation of the 10 optimi­
swarm location (x 0 , y0 ) is acquired by Eq. (14). zation results are recorded. It can be found that MFOA outperforms
FOA in that the optimization results improve for at least one order of
x 0 = rand (LR) magnitude. For the Rastrigin function ( f2 ), MFOA finds its global op­
y0 = rand (LR) (14) timal value directly. The smaller standard deviation of MFOA also re­
flects its better optimization reliability.
Step 2: Initialization. The random location (xi , yi ) and distance for Fig. 2 illustrates the convergence curve of MFOA and FOA for the
foraging of an individual fruit fly are given by smell. The proposed step- three benchmark functions. It shows that in the iterative process of all
size adjustment coefficient (k ) based on Logistic function is introduced. three functions, MFOA converges slower at beginning, but thereafter, it
picks up faster than FOA. The dynamic step-size search strategy based
xi = x 0 + k· rand (FR) on Logistic function enables MFOA to search in a wider range initially
yi = y0 + k· rand (FR) (15) to avoid falling into local optimum, which indicates that the mod­
ification works. It is worth mentioning that although the above mod­
1.00 ification improves the convergence accuracy, but it extends the search
maxstep = 1 time. Nevertheless, there is no significant difference in the number of
maxgen = 20 iterations required to achieve optimum between MFOA and FOA. Spe­
cifically, 53, 25, and 54 iterations are required for MFOA to obtain
0.75
optimal results of the three benchmark functions, respectively. In
general, the performance of MFOA is much better than that of the
original FOA, and this confirms the effectiveness of the proposed
Step-size

Decrease Decrease Remain modification.


0.50
slowly quickly stable
3.3.4. MFOA-SVM combination prediction model
As discussed in Section 3.3.1, SVM performance is mainly affected
0.25
by the approximation degree of two key hyper-parameters, i.e., the
penalty factor (C ) and the bandwidth of the RBF ( ). C represents the
tolerance degree of error, and improper value of this parameter will
0.00 lead to poor generalization ability. determines the distribution of data
0 5 10 15 20 mapping to high-dimensional space, which affects the number of sup­
Iterations port vectors. Therefore, C and have to do with the running speed and
prediction accuracy of SVM, namely their values determine the quality
Fig. 1. Mapping curve based on Logistic function. of the prediction. To make a better use of the capacity of SVM, these

5
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

Table 2
Basic information of three benchmark functions.
Function Expression Search range Optimal value Peak

Sphere f1 = x 2 + y 2 [−3.00, 3.00] 0 unimodal


Rastrigin f2 = 20 + x 2 10 cos(2 x ) + y 2 10 cos(2 y ) [−5.12, 5.12] 0 multimodal
Ackley 0.2 0.5(x 2 + y 2 ) [−5.00, 5.00] 0 multimodal
f3 = 20 20e e 0.5cos (2 x) + e0.5 cos(2 y)

Table 3 these indexes, RMSE represents the squared mean deviation of the
Optimization results of three benchmark functions. predicted value from the measured value. MAE is the average of ab­
Function FOA MFOA
solute errors from the center. MAPE is a measure of the accuracy error
in percentage form. R2 is a common index to evaluate the goodness of
Mean Standard deviation Mean Standard deviation fit of the model. These are cost-type indexes other than R2, meaning
that the smaller the index value is, the better the model performs. In
Sphere 5.03E-04 6.02E-05 2.04E-43 1.24E-42
Rastrigin 2.93E-01 3.12E-01 0.00 0.00
addition, an improved comprehensive index (ICI), derived from RMSE,
Ackley 9.45E-02 8.79E-03 3.55E-15 4.37E-15 MAE, MAPE and R2, is proposed to evaluate the overall performance of
prediction models, which is given by Eq. (22).

p q
two SVM hyper-parameters are dynamically adjusted using the MFOA 1 Pi Pmin, i Pmax, j Pj
ICI = + , 0 ICI 1
technique presented in Section 3.3.2, and 5-fold cross validation using p+q Pmax, i Pmin, i Pmax , j Pmin, j
i=1 j=1
the training samples. When the optimized hyper-parameter values are
generated, a novel combination prediction model based on SVM and (22)
MFOA is established. where p and q are the number of cost-type and profit-type indexes re­
spectively, and Pi is the i-th evaluation index. The smaller ICI value
3.4. Performance measurement criteria indicates better performance of a model.

The following performance measures, root mean square error


3.5. Construction procedure of combination deformation prediction model
(RMSE), mean absolute error (MAE), mean absolute percentage error
considering quantitative evaluation and hysteresis correction
(MAPE), and coefficient of determination (R2) are adopted to compre­
hensively evaluate the performance of deformation prediction models.
On the basis of the aforementioned theories and methods, the
The mathematical expressions of these indexes are shown in Eqs.
combination deformation prediction model considering quantitative
(18)–(21).
evaluation and hysteresis correction is proposed. The modeling proce­
n 2 dure of the combination prediction model is shown in Fig. 3, and also
1
RMSE = ym(i) yp(i ) summarized below.
n (18)
i=1 Step 1: The conventional factors influencing dam deformation are
n selected as input variables by using the HST model.
1
MAE = |ym(i ) yp(i) | Step 2: The hysteresis of dam deformation to upstream water level
n (19)
i=1 and air temperature, namely the newly added input variable, is de­
termined by the hysteresis quantification algorithm (in Section 3.1).
100%
n ym(i ) yp(i) Step 3: The SEV value of each normalized input variable, except the
MAPE =
n i=1 ym(i) (20) new hysteresis variables, is calculated by the synthetic evaluation
method (in Section 3.2).
n 2
Step 4: The corrected training samples, including input and output
ym(i) yp(i) variables, are imported into the SVM, and the presented MFOA is em­
i=1
R2 = 1 n ployed to optimize the SVM hyper-parameters (in Section 3.3). Also, the
2
(ym(i) ym ) derived ICI (in Section 3.4) is taken as the fitness function of MFOA.
(21)
i=1
Step 5: The combination deformation prediction model considering
where n is the total number of data samples, and denote the i-th ym(i ) yp(i ) quantitative evaluation and hysteresis correction is constructed on ac­
measured and predicted value, and ym is the mean of ym values. Among count of the optimized SVM hyper-parameter values.

0.04
FOA 18 FOA 6 FOA
MFOA MFOA MFOA
0.03 15 5
Fitness value

Fitness value

Fitness value

12 4
0.02
9 3

6 2
0.01
3 1

0.00 0 0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Iterations (a) Iterations (b) Iterations (c)
Fig. 2. Optimization results of FOA and MFOA for three benchmark functions: (a) Sphere function; (b) Rastrigin function; (c) Ackley function.

6
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

Hysteresis quantification Training samples (input


HST statistical model
algorithm and output variables)
Decomposition Reconstruction

Eleven conventional Four new hysteresis Parameter initialization


influencing factors factors of SVM and MFOA

H, H2, H3, Ts1, Ts2, Ts3,


Hh, Hh2, Hh3, Th
Tc1, Tc2, Tc3, θ, lnθ
Smell searching process

MFOA
Synthetic evaluation
Variable normalization
method Vision searching process

Preprocessed dam Yes


deformation database gen < maxgen

Normalization No
Dam Validation samples (input Performance evaluation Obtain the SVM with
deformation and output variables) optimal hyper-parameters
Fig. 3. Flowchart of model formation.

Step 6: The validation samples are imported to test the performance horizontal deformation both along flow direction (i.e., X direction) and
of the trained MFOA-SVM model, and there are several indexes (in vertical flow direction (i.e., Y direction). The dam deformation, water
Section 3.4) that can be used to evaluate the fitting and extrapolation level and air temperature are measured and stored automatically,
performance of the proposed dam deformation prediction model (i.e., which makes the modeling of dam behavior feasible.
SEV-MFOA-SVM). The deformation series of three monitoring points PL17Y (to the left
bank is positive, to the right bank is negative), PL29X and PL17X (to the
downstream is positive, to the upstream is negative) in the powerhouse
4. Case study on the right bank are selected as test cases in this study. The monitoring
data of deformation and environmental factors (i.e., water level and air
The studied dam project is located on the Yangtze River in Yichang, temperature) are illustrated in Fig. 5. These data were generated by the
Hubei Province, China, as shown in Fig. 4. It is a 47-meter-high, 2595- automatic monitoring system and also subsequently proofread by the
meter-long concrete dam, mainly composed of navigation lock, power facility engineers. It should be noted that PL17Y is taken as a re­
station, discharging sluice, and scouring sluice. The project serves in presentative example in this section, and the other two monitoring
generating electricity and improving channel conditions. The con­ points are used for further model validation in Section 5.3. The PL17Y
struction of the dam started in 1970 and completed in 1988. To ensure deformation series corresponds to a period between 5 January 1994
its safety operation, an automatic safety monitoring system was de­ and 24 April 2014, resulting in a total of 307 observations. 236 groups
ployed on the dam surface and inside of its main structures. The system of data from 5 January 1994 to 25 April 2013 are assigned as training
consists of a number of instruments that measure quantities such as samples (left to the vertical line in Fig. 5), and 71 groups of data from
water level, air temperature, horizontal deformation, seepage, etc. 30 April 2013 to 24 April 2014 are assigned as validation samples (right
Plumb lines (PL) and inverted plumb lines (IP) are used to monitor

Fig. 4. Basic information of the studied dam project.

7
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

67

Upstream water level (m)


66

65

64

63

30
)
Air temperature (

20

10

0
Deformation (mm)

-4

-8

-12
PL17Y PL17X PL29X
1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013
Time (year)
Fig. 5. Hydrographs of measured deformation and environment variables.

to the vertical line in Fig. 5). The cubic Hermite interpolator [59] is monitoring date to the initial monitoring date, and = t 100 . For the
used to homogenize the training samples, and the interpolation results dam project studied, m and n are both set as 3. As a result, the con­
with equal intervals (one data per month) are obtained. Validation
2 t 4 t 6 t
ventional influencing factors include H , H 2 , H 3 , sin 365 , sin 365 , sin 365 ,
samples with unequal intervals are not preprocessed except for nor­ 2 t 4 t 6 t 2 t 4 t
cos 365 , cos 365 , cos 365 , and ln . For convenience, sin 365 , sin 365 , sin 365 ,
6 t

malization. 2 t 4 t 6 t
cos 365
,
cos and cos
365
are represented by Ts1, Ts2 , Ts3, Tc1, Tc2 and Tc3 .
365
To improve the quantification accuracy, the original deformation
4.1. Selection of influencing factors series is interpolated into a new series with a time interval of 5 days
(assume 360 days in a year). On the basis of the hysteresis quantifica­
The HST model is the most common statistical model for dam de­ tion algorithm, the hysteresis calculation results of PL17Y deformation
formation prediction [21]. Most of the previous mathematical models relative to both upstream water level and air temperature are shown in
for dam safety monitoring were developed based on the HST model. Fig. 6. It is observed that the hysteresis time (Th ) of PL17Y deformation
Thus, the conventional influencing factors can be selected by the model. to air temperature is 10 days, and the hysteresis time (Hh ) of the de­
However, the HST model does not contain hysteresis of dam deforma­ formation to upstream water level is also 10 (i.e., 360 350 = 10 ) days
tion relative to water level and air temperature; also, it simulates the according to the proximity principle. Referring to the mathematical
temperature effect via harmonics without reflecting the fluctuation of formula of H , the hysteresis hydrostatic component ( Hh ) is expressed
actual temperature. It is necessary to consider the actual hysteresis 3

factors by the hysteresis quantification algorithm discussed in Section as i


di H10 , while the hysteresis temperature component ( Th ) is ex­
i=1
3.1. The input variables of the proposed combination prediction model pressed as eT10 , where di and e denote unknown coefficients. Therefore,
are composed of conventional influencing factors and the four new the new hysteresis factors are H10 , H10
2
, H10
3
and T10 .
hysteresis factors. The HST model in Eq. (23) is a linear regression model, which can
According to the HST model, the dam deformation ( ) at any point more intuitively describe the contribution of each factor. SR is adopted
can be divided into the hydrostatic component ( H ), temperature to solve Eq. (23) to prove the necessity of the new hysteresis factors.
component ( T ), and aging component ( ), which is given by Eq. (23). Two HST models can be obtained by SR, one containing only conven­
tional influencing factors, and the other containing all influencing
= H + T + factors including hysteresis factors. The fitting and prediction results of
n m
= a0 + ai H i + b1i sin
2 it
+ b2i cos
2 it
+ c1 + c2 ln PL17Y deformation series by the two models are compared in Fig. 7. A
i=1 i=1 365 365 new assessment index, relative error (RE), is added to visually display
(23) the prediction error in the chart, which is expressed as

where a0 is the constant term, ai , b1i , b2i , c1 and c2 represent statistical RE = | ym(i ) yp(i) ym(i ) |. It can be found that the model fitting effect is
coefficients, H is the upstream water level, n is the coefficient of dam little improved by the addition of hysteresis factors, while the model
type, m is the number of cycles, t denotes the cumulative days from the

8
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

Maximum Table 4
1.0
Assessment of fitting and prediction performance of stepwise regression for
Th Hh PL17Y deformation with and without hysteresis quantification.
0.8
Indexes SR with hysteresis quantification SR without hysteresis quantification
Cosine similarity

0.6 Fitting Prediction Fitting Prediction


Maximum
RMSE/mm 0.398 0.352 0.405 0.388
0.4
MAE/mm 0.311 0.281 0.313 0.315
MAPE 0.186 0.103 0.186 0.114
0.2 R2 0.855 0.897 0.850 0.875

0.0
set as 0.5. Statistics for the normalized PL17Y deformation series are
10 100 200 300 350 summarized in Table 5, and the synthetic evaluation results are shown
in Table 6. The resulted SEV determines the contribution of each factor
Hysteresis time (d)
to dam deformation. Each normalized factor series is multiplied by the
Fig. 6. The hysteresis time of PL17Y deformation relative to upstream water corresponding SEV to acquire the preprocessed factor series {SEV1·H ,
level and air temperature. SEV2·H 2 , SEV3· H 3 , SEV4· Ts1, SEV5·Ts2 , SEV6·Ts3 , SEV7·Tc1, SEV8· Tc 2 ,
SEV9· Tc3 , SEV10· , SEV11·ln } . As for the four hysteresis factors, they are
prediction performance is indeed improved to some extent. Compared not weighted (equivalent to multiplying by one) to avoid weakening
with the SR without hysteresis quantification, the SR with hysteresis their contribution. In other words, they are still {H10 , H10 2
, H103
, T10} .
quantification can grasp the fluctuation of PL17Y deformation series, as Therefore, the final input variables of the combination deformation
shown in Fig. 7(b). The performance assessment results in Table 4 also prediction model considering quantitative evaluation and hysteresis
confirm the above observation, indicating that it is feasible to introduce correction are {SEV1·H , SEV2·H 2 , SEV3· H 3, SEV4· Ts1, SEV5·Ts2 , SEV6·Ts3 ,
four hysteresis factors {H10 , H10
2 3
, H10 , T10} as model input variables. SEV7·Tc1, SEV8· Tc 2 , SEV9· Tc3 , SEV10· , SEV11·ln , H10 , H10
2
, H10
3
, T10} .

4.2. Synthetic evaluation of influencing factors 4.3. Schematic design of simulation analysis

The proposed combination prediction model considering hysteresis The usefulness of hysteresis quantification is verified by the HST
quantification is applied for PL17Y deformation prediction, and be­ model in Section 4.1, and simulation experiments are necessary to
cause four hysteresis factors are desired as discussed previously, thus prove the improvement effect of hyper-parameter optimization and
the following hysteresis input variables are included: {H , H 2 , H 3 , Ts1, quantitative evaluation. The proposed combination deformation pre­
Ts2 , Ts3 , Tc1, Tc2 , Tc3 , , ln , H10 , H10
2
, H10
3
, T10} . The 11 conventional in­ diction model considering quantitative evaluation and hysteresis cor­
fluencing factors are weighted by the synthetic evaluation method rection is referred to as SEV-MFOA-SVM for short. Five other de­
presented in Section 3.2, where the weight allocation coefficient ( ) is formation prediction models are established to compare with SEV-

0
Deformation (mm)

-1

-2

-3

-4

-5 Measured value SR with hysteresis quantification SR without hysteresis quantification

1993 1995 1997 1999 2001 2003 2005 2007 2009 2011
Time (year) (a)
0 300
Measured value
SR with hysteresis
-1 250
quantification
Deformation (mm)

Relative error (%)

SR without hysteresis
-2 200
quantification
-3 150 Relative error of SR with
hysteresis quantification
-4 100 Relative error of SR without
hysteresis quantification
-5 50

-6 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (b)
Fig. 7. Fitting and prediction performance comparisons of stepwise regression for PL17Y deformation with and without hysteresis quantification: (a) fitting; (b)
prediction.

9
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

Table 5
Data statistics for normalized PL17Y deformation series with 11 conventional influencing factors normalized.
Statistics H H2 H3 Ts1 Ts2 Ts3 Tc1 Tc2 Tc3 θ lnθ δ

Mean 0.69 0.68 0.68 0.51 0.50 0.50 0.50 0.50 0.50 0.50 0.85 0.51
Variance 0.06 0.06 0.06 0.13 0.13 0.13 0.13 0.13 0.13 0.08 0.02 0.04
StdDev 0.24 0.24 0.24 0.35 0.35 0.35 0.35 0.36 0.36 0.29 0.16 0.20

* StdDev stands for standard deviation.

Table 6 MFOA-SVM, which are the grid search-optimized SVM (GRID-SVM),


Synthetic evaluation of 11 conventional influencing factors of PL17Y. FOA-optimized SVM (FOA-SVM), MFOA-optimized SVM (MFOA-SVM),
Value H H2 H3 Ts1 Ts2 Ts3 Tc1 Tc2 Tc3 θ lnθ
grid search-optimized SVM considering quantitative evaluation (SEV-
GRID-SVM) and FOA-optimized SVM considering quantitative evalua­
CVj 0.35 0.35 0.36 0.70 0.70 0.71 0.70 0.71 0.71 0.58 0.18 tion (SEV-FOA-SVM). To make it fair, the six models all consider hys­
ICj 0.07 0.07 0.07 0.11 0.11 0.11 0.07 0.07 0.07 0.13 0.13 teresis quantification, that is, their input variables are unified as {H , H 2 ,
ωj 0.05 0.05 0.05 0.14 0.14 0.14 0.08 0.09 0.09 0.13 0.04
H 3 , Ts1, Ts2 , Ts3 , Tc1, Tc2 , Tc3 , , ln , H10 , H10
2
, H10
3
, T10} , and the output
RCj 0.09 0.09 0.09 0.07 0.08 0.09 0.16 0.08 0.08 0.09 0.09
SEVj 0.07 0.07 0.07 0.10 0.11 0.12 0.12 0.08 0.08 0.11 0.06 variable is dam deformation. To ensure the rationality of comparison
experiments, the hyper-parameters of the models optimized with FOA
and MFOA are set as follows: C [2 10 , 210], [2 15 , 25], maxgen = 20

0
Deformation (mm)

-1

-2

-3

-4
Measured value SEV-GRID-SVM GRID-SVM
-5

1993 1995 1997 1999 2001 2003 2005 2007 2009 2011
Time (year) (a)
1

0
Deformation (mm)

-1

-2

-3

-4
Measured value SEV-FOA-SVM FOA-SVM
-5

1993 1995 1997 1999 2001 2003 2005 2007 2009 2011
Time (year) (b)
1

0
Deformation (mm)

-1

-2

-3

-4
Measured value SEV-MFOA-SVM MFOA-SVM
-5

1993 1995 1997 1999 2001 2003 2005 2007 2009 2011
Time (year) (c)
Fig. 8. Fitting performance comparison of six prediction models for PL17Y deformation: (a) GRID-SVM and SEV-GRID-SVM; (b) FOA-SVM and SEV-FOA-SVM; (c)
MFOA-SVM and SEV-MFOA-SVM. The green-shaded areas represent the data points prone to large prediction errors.

10
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

and sizepop = 20 , and the hyper-parameters of the models optimized optimization, but Table 7 makes up for it. It can be known from Table 7
with grid search algorithm are set as C {2 10 , 2 9, , 29, 210} and that MFOA-SVM has the best fitting effect among the GRID-SVM, FOA-
{2 15 , 2 14, , 24 , 25} . Also, the maximum search step-size of C and SVM and MFOA-SVM models. Similarly, the fitting effect of SEV-MFOA-
are 10 and 1 in MFOA, respectively. It should be mentioned that all the SVM in the three models of SEV-GRID-SVM, SEV-FOA-SVM and SEV-
deformation prediction results of each model are the average values MFOA-SVM is also the optimal. Also, there is little difference between
taken through 10 runs. Accordingly, the scheme of simulation analysis the fitting effect of models optimized by grid search algorithm and FOA,
is arranged as follows. neither of which is better than the MFOA-optimized model. These re­
Scheme 1: Verification of hyper-parameter optimization effect. sults show that modifying FOA can be an effective approach in im­
Taking PL17Y deformation for example, the effect of hyper-parameter proving model-fitting performance. In addition, quantitative evaluation
optimization is proved by comparing the fitting and prediction perfor­ and hyper-parameter optimization are compared for the improvement
mance of GRID-SVM, FOA-SVM and MFOA-SVM. The same is true for increment of fitting performance in Table 7. It is found that compared
SEV-GRID-SVM, SEV-FOA-SVM and SEV-MFOA-SVM. The above two with hyper-parameter optimization, quantitative evaluation can im­
groups of comparison experiments are set up to avoid contingency, see prove model fitting performance more dramatically.
Sections 5.1 and 5.2.
Scheme 2: Verification of quantitative evaluation effect. As an 5.2. Prediction performance evaluation
example of PL17Y deformation, the effect of quantitative evaluation is
verified by comparing the fitting and prediction performance of models The six deformation prediction models trained in Section 5.1 were
with and without quantitative evaluation. Three groups of comparison tested with the validation samples. Fig. 9 displays the prediction com­
experiments are conducted, which are GRID-SVM versus SEV-GRID- parison results of the six models, and the assessment results of each
SVM, FOA-SVM versus SEV-FOA-SVM, and MFOA-SVM versus SEV- model are shown in Table 8. On account of the sudden change of the
MFOA-SVM, see Sections 5.1 and 5.2. upstream water level (see Fig. 5), the models without quantitative
Scheme 3: Verification of model generalization capability. The evaluation (i.e., GRID-SVM and FOA-SVM) have large prediction error
flexibility of the proposed SEV-MFOA-SVM model is proved by the from 30 July 2013 to 5 October 2013, as shown in the green-shaded
deformation prediction results of two monitoring points (i.e., PL29X areas of Fig. 9. The difference lies in that the three models considering
and PL17X) except PL17Y, as shown in Section 5.3. quantitative evaluation all show excellent prediction effect without too
much error. It shows that quantitative evaluation can reduce the in­
fluence of sudden value change of input variables on prediction. A ro­
5. Results and discussion
bust deformation prediction model is desired to better handle abnormal
changes in environmental factors. It can be seen from the indexes in
5.1. Fitting performance evaluation
Table 8 that models considering quantitative evaluation have better
predictions, and the SEV-MFOA-SVM model has the highest prediction
Fig. 8(a)–(c) show the training comparison results of each model
accuracy. The relative error in Fig. 9 also reflects the fact that the
(i.e., GRID-SVM, FOA-SVM and MFOA-SVM) with and without quan­
prediction error of the models considering quantitative evaluation is
titative evaluation. The fitting performance of six models is assessed
reduced. As such, integrating quantitative evaluation into deformation
with measurement indexes, as summarized in Table 7. It can be seen
prediction model can bring some improvements to the accuracy and
from the three subgraphs in Fig. 8 that the fitting curves of models
reliability of prediction model.
considering quantitative evaluation are relatively smooth, while the
Table 8 clearly displays the improvement effect of hyper-parameter
fitting curves of the corresponding models without quantitative eva­
optimization. Whether in the three models considering quantitative
luation have many abrupt changes. Among all the input variables, up­
evaluation or in the three models without considering quantitative
stream water level and air temperature vary significantly, which can
evaluation, the indexes of the MFOA-optimized models (i.e., MFOA-
affect model reliability. It appears that the model with quantitative
SVM and SEV-MFOA-SVM) are optimal. Compared with the models
evaluation is more reliable. In Table 7, the indexes of models con­
optimized by grid search algorithm, the of FOA-optimized models do
sidering quantitative evaluation are all better than those of the corre­
not show noticeable improvement. Thus, the modified FOA contributes
sponding models without quantitative evaluation, indicating the fitting
to enhance the prediction performance of the deformation prediction
accuracy of the model with quantitative evaluation is better. The SEV-
model. In terms of stability, among the three models without quanti­
MFOA-SVM model has the best fitting performance among the six
tative evaluation, only MFOA-SVM has no significant error in de­
models, and its fitting curve almost coincides with the measured curve,
formation prediction from 30 July 2013 to 5 October 2013, see
see Fig. 8(c). It is worth mentioning that the fitting errors of the other
Fig. 9(c). MFOA can improve the accuracy and reliability of deforma­
five models for the deformation monitoring data on 29 September 2006
tion prediction model at the same time. Besides, it can be discovered
are all large, except for SEV-MFOA-SVM, as shown in the green-shaded
from Table 8 that quantitative evaluation can improve model prediction
areas of Fig. 8. The mutation resistance and robustness of the SEV-
performance more significantly than hyper-parameter optimization.
MFOA-SVM model is demonstrated in this case.
Fig. 8 cannot directly show the effect of hyper-parameter
5.3. Further validation and discussion
Table 7
PL29X and PL17X deformation series (see Fig. 5), which show de­
Assessment of fitting performance of six prediction models for PL17Y de­
formation.
formation in a different direction than PL17Y, are selected to further
compare the prediction performance of the proposed SEV-MFOA-SVM
Indexes GRID-SVM FOA-SVM MFOA-SVM SEV- SEV- SEV- model with other models. According to the procedure in Section 4.1, all
GRID- FOA- MFOA-
SVM SVM SVM
the factors influencing the dam deformation of the two monitoring
points are determined. The conventional influencing factors of both are
RMSE/mm 0.453 0.456 0.398 0.403 0.395 0.390 the same as those of PL17Y, hysteresis factors of PL29X deformation are
MAE/mm 0.364 0.370 0.321 0.324 0.322 0.318 2
{H165, H165 3
, H165 , T135} , while hysteresis factors of PL17X deformation are
MAPE 0.230 0.235 0.188 0.191 0.186 0.183 2
{H30 , H30 3
, H30 , T55} . After specifying the input variables of six models, the
R2 0.902 0.901 0.928 0.928 0.931 0.932
ICI 0.931 1.000 0.105 0.148 0.063 0.000 three models considering quantitative evaluation (i.e., SEV-GRID-SVM,
SEV-FOA-SVM and SEV-MFOA-SVM) should utilize SEV to weight the
* Bold is better. conventional influencing factors based on the steps in Section 4.2. The

11
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

1 300
Measured value
0 SEV-GRID-SVM
250
Deformation (mm) GRID-SVM
-1

Relative error (%)


200 Relative error of
-2
SEV-GRID-SVM
-3 150 Relative error of
-4 GRID-SVM
100
-5
50
-6
-7 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (a)

0 300
Measured value
SEV-FOA-SVM
-1 250
FOA-SVM
Deformation (mm)

Relative error (%)


-2 200 Relative error of
SEV-FOA-SVM
-3 150 Relative error of
FOA-SVM
-4 100

-5 50

-6 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (b)

0 300
Measured value
SEV-MFOA-SVM
-1 250
MFOA-SVM
Deformation (mm)

Relative error (%)

-2 200 Relative error of


SEV-MFOA-SVM
-3 150 Relative error of
MFOA-SVM
-4 100

-5 50

-6 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (c)
Fig. 9. Prediction performance comparison of six prediction models for PL17Y deformation: (a) GRID-SVM and SEV-GRID-SVM; (b) FOA-SVM and SEV-FOA-SVM; (c)
MFOA-SVM and SEV-MFOA-SVM. The green-shaded areas represent the data points prone to large prediction errors.

Table 8 six deformation prediction models are trained with training samples
Assessment of prediction performance of six prediction models for PL17Y de­ and then tested with validation samples.
formation. Fig. 10 exhibits the PL29X deformation prediction results of the six
Indexes GRID-SVM FOA-SVM MFOA-SVM SEV- SEV- SEV- models, and the corresponding performance assessment is shown in
GRID- FOA- MFOA- Table 9. From Fig. 10, the three models without quantitative evaluation
SVM SVM SVM are not stable, and there is a large prediction error between 20 No­
vember 2013 and 5 March 2014, as shown in the green-shaded areas.
RMSE/mm 0.568 0.544 0.367 0.437 0.440 0.293
MAE/mm 0.412 0.428 0.305 0.366 0.369 0.244 After introducing the quantitative evaluation, the stability of the three
MAPE 0.163 0.170 0.116 0.137 0.138 0.085 models is improved. According to the relative error in Fig. 10, SEV-
R2 0.880 0.873 0.956 0.949 0.949 0.972 MFOA-SVM has the best prediction performance. Moreover, it is clear
ICI 0.940 0.979 0.283 0.509 0.516 0.000 from Table 9 that MFOA-SVM has a big lead over both GRID-SVM and
FOA-SVM in regards to the performance indexes, which shows that the
* Bold is better.
improvement of model prediction performance can be partly attributed
to the hyper-parameter optimization. In summary, the proposed SEV-

12
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

-4 300
Measured value
-6 SEV-GRID-SVM
250
GRID-SVM
Deformation (mm) -8

Relative error (%)


200 Relative error of
-10 SEV-GRID-SVM
150 Relative error of
-12 GRID-SVM
100
-14

-16 50

-18 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (a)

-4 300
Measured value
-6 SEV-FOA-SVM
250
FOA-SVM
Deformation (mm)

-8

Relative error (%)


200 Relative error of
-10 SEV-FOA-SVM
150 Relative error of
-12 FOA-SVM
100
-14

-16 50

-18 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (b)

-4 300
Measured value
-6 SEV-MFOA-SVM
250
MFOA-SVM
Deformation (mm)

-8
Relative error (%)

200 Relative error of


-10 SEV-MFOA-SVM
150 Relative error of
-12 MFOA-SVM
100
-14

-16 50

-18 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (c)
Fig. 10. Prediction performance comparison of six prediction models for PL29X deformation: (a) GRID-SVM and SEV-GRID-SVM; (b) FOA-SVM and SEV-FOA-SVM;
(c) MFOA-SVM and SEV-MFOA-SVM. The green-shaded areas represent the data points prone to large prediction errors.

Table 9 MFOA-SVM model has superior adaptability to PL29X deformation, and


Assessment of prediction performance of six prediction models for PL29X de­ the combination of quantitative evaluation and hyper-parameter opti­
formation. mization can enhance the precision and robustness of the original de­
Indexes GRID-SVM FOA-SVM MFOA-SVM SEV- SEV- SEV- formation prediction model.
GRID- FOA- MFOA- The PL17X deformation prediction results of the six models are
SVM SVM SVM shown in Fig. 11, and their performance assessment is displayed in
Table 10. Obviously, unlike the deformation prediction effect of PL17Y
RMSE/mm 2.048 2.113 1.335 0.843 1.314 0.663
MAE/mm 1.543 1.569 1.108 0.633 1.092 0.526 and PL29X, the deformation prediction curve of PL17X is relatively
MAPE 0.147 0.149 0.105 0.066 0.107 0.055 smooth without significant fluctuations as shown in Fig. 11. Compared
R2 0.674 0.660 0.864 0.956 0.865 0.973 with the models without quantitative evaluation, the correction of
ICI 0.966 1.000 0.475 0.100 0.472 0.000 quantitative evaluation to prediction results is almost unnoticeable, as
can be observed in the relative error. In Table 10, the performance
* Bold is better.
assessment results of the six models are close, and SEV-MFOA-SVM do
not show any significant performance advantage. It indicates that

13
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

1 300
Measured value
0 SEV-GRID-SVM
250
Deformation (mm) GRID-SVM
-1

Relative error (%)


200 Relative error of
-2
SEV-GRID-SVM
-3 150 Relative error of
-4 GRID-SVM
100
-5
50
-6
-7 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (a)

1 300
Measured value
0 SEV-FOA-SVM
250
FOA-SVM
Deformation (mm)

-1

Relative error (%)


200 Relative error of
-2
SEV-FOA-SVM
-3 150 Relative error of
-4 FOA-SVM
100
-5
50
-6
-7 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (b)

1 300
Measured value
0 SEV-MFOA-SVM
250
MFOA-SVM
Deformation (mm)

-1
Relative error (%)

200 Relative error of


-2
SEV-MFOA-SVM
-3 150 Relative error of
-4 MFOA-SVM
100
-5
50
-6
-7 0
2013/4/25 2013/6/25 2013/8/25 2013/10/25 2013/12/25 2014/2/25 2014/4/25
Time (year/month/day) (c)
Fig. 11. Prediction performance comparison of six prediction models for PL17X deformation: (a) GRID-SVM and SEV-GRID-SVM; (b) FOA-SVM and SEV-FOA-SVM;
(c) MFOA-SVM and SEV-MFOA-SVM.

Table 10 hyper-parameter optimization, like quantitative evaluation, has no ef­


Assessment of prediction performance of six prediction models for PL17X de­ fect on PL17X deformation prediction. From the comparison with
formation. PL17Y and PL29X deformation prediction, if the prediction results of
Indexes GRID-SVM FOA-SVM MFOA-SVM SEV- SEV- SEV- the original model are good, the improvement effect of quantitative
GRID- FOA- MFOA- evaluation and hyper-parameter optimization would not be significant.
SVM SVM SVM On the contrary, for the original model with poor prediction, both
measures can greatly improve the prediction precision and reliability of
RMSE/mm 0.451 0.447 0.405 0.450 0.441 0.403
MAE/mm 0.386 0.388 0.338 0.409 0.399 0.346 the model. In short, the two measures proposed in this study, namely
MAPE 0.308 0.317 0.244 0.320 0.300 0.278 the quantitative evaluation and hyper-parameter optimization in the
R2 0.973 0.973 0.977 0.979 0.978 0.980 SEV-MFOA-SVM model, will produce different improvement effects at
ICI 0.880 0.895 0.118 0.781 0.668 0.140
various monitoring points.
* Bold is better.

14
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

6. Conclusions and future work [2] J. Jia, Dam construction in China: A sixty-year review, China Water & Power Press,
Beijing, 2013.
[3] F. Kang, J. Li, J. Dai, Prediction of long-term temperature effect in structural health
An optimized combination prediction model for concrete dam de­ monitoring of concrete dams using support vector machines with Jaya optimizer
formation considering quantitative evaluation and hysteresis correc­ and salp swarm algorithms, Adv. Eng. Softw. 131 (2019) 60–76.
tion, abbreviated as SEV-MFOA-SVM, is proposed in this study. The [4] C. Shao, C. Gu, M. Yang, et al., A novel model of dam displacement based on panel
data, Struct. Control Health Monitor. 25 (1) (2018) e2037.
conventional deformation prediction model based on SVM is improved [5] C. Xu, D. Yue, C. Deng, Hybrid GA/SIMPLS as alternative regression model in dam
in three aspects, that is, phase correction, hyper-parameter optimiza­ deformation analysis, Eng. Appl. Artif. Intell. 25 (3) (2012) 468–475.
tion, and synthetic evaluation. Taking a real-world dam project as a [6] F. Salazar, R. Morán, M.Á. Toledo, et al., Data-based models for the prediction of
dam behaviour: A review and some methodological considerations, Arch. Comput.
case study, the effectiveness of the presented hysteresis quantization Methods Eng. 24 (1) (2017) 1–21.
algorithm, MFOA and synthetic evaluation method is verified through [7] Y. Yu, X. Liu, E. Wang, et al., Dam safety evaluation based on multiple linear re­
simulation comparison experiments. The results show that the SEV- gression and numerical simulation, Rock Mech. Rock Eng. 51 (8) (2018)
2451–2467.
MFOA-SVM model proposed is superior to other methods in modeling
[8] S. Chen, C. Gu, C. Lin, et al., Safety monitoring model of a super-high concrete dam
and predicting dam behavior, and thus a valuable solution for similar by using RBF neural network coupled with kernel principal component analysis,
applications. The main conclusions of this work are summarized as Mathematical Probl. Eng. (2018) 1–13.
follows. [9] M. Tatin, M. Briffaut, F. Dufour, et al., Thermal displacements of concrete dams:
Accounting for water temperature in statistical models, Eng. Struct. 91 (2015)
26–39.
(1) Four new hysteresis factors {Hh, Hh2, Hh3, Th} , besides the conven­ [10] P. Milillo, D. Perissin, J.T. Salzer, et al., Monitoring dam structural health from
tional influencing factors, are introduced by the hysteresis quanti­ space: Insights from novel InSAR techniques and multi-parametric modeling ap­
plied to the Pertusillo dam Basilicata, Italy, Int. J. Appl. Earth Obs. Geoinf. 52
zation algorithm. Simulating the influence of preceding upstream (2016) 221–229.
water level and air temperature on the current deformation, the [11] B. Wei, D. Yuan, Z. Xu, et al., Modified hybrid forecast model considering chaotic
monitoring model can accurately capture the evolution trend of residual errors for dam deformation, Struct. Control Health Monitor. 25 (8) (2018)
e2188.
dam deformation. [12] S.W. Wang, C.S. Gu, T.F. Bao, Observed displacement data-based identification
(2) The synthetic evaluation method quantifies the contribution of each method of deformation time-varying effect of high concrete dams, Sci. China
influencing factor to dam deformation. This measure really reflects Technolog. Sci. 61 (6) (2018) 906–915.
[13] B. Stojanovic, M. Milivojevic, N. Milivojevic, et al., A self-tuning system for dam
the different influence of each factor on dam deformation, which behavior modeling based on evolving artificial neural networks, Adv. Eng. Softw.
can improve the accuracy and robustness of the original deforma­ 97 (2016) 85–95.
tion prediction model. [14] C.Y. Kao, C.H. Loh, Monitoring of long-term static deformation data of Fei-Tsui arch
dam using artificial neural network-based approaches, Struct. Control Health
(3) Compared with the original FOA, the convergence accuracy of the
Monitor. 20 (3) (2013) 282–303.
proposed MFOA is enhanced. MFOA makes great contribution to­ [15] V. Ranković, N. Grujović, D. Divac, et al., Development of support vector regression
ward taking the full advantage of SVM's nonlinear mapping cap­ identification model for prediction of dam structural behaviour, Struct. Saf. 48
ability. Also, it is found that MFOA can improve the reliability of (2014) 33–39.
[16] H. Su, Z. Chen, Z. Wen, Performance improvement method of support vector ma­
existing prediction models in practice. chine-based model monitoring dam safety, Struct. Control Health Monitor. 23 (2)
(4) The SEV-MFOA-SVM model is a successful combination of hyster­ (2016) 252–266.
esis correction, hyper-parameter optimization and synthetic eva­ [17] H. Su, X. Li, B. Yang, et al., Wavelet support vector machine-based prediction model
of dam deformation, Mech. Syst. Sig. Process. 110 (2018) 412–427.
luation, which has excellent learning and generalization capability. [18] F. Kang, J. Liu, J. Li, et al., Concrete dam deformation prediction model for health
It should be noted that the above three measures could produce monitoring based on extreme learning machine, Struct. Control Health Monitor. 24
different improvement effects at different monitoring locations. (10) (2017) e1997.
[19] K.T.T. Bui, D.T. Bui, J. Zou, et al., A novel hybrid artificial intelligent approach
According to our simulations, improvements are more remarkable based on neural fuzzy inference model and particle swarm optimization for hor­
where traditional models have poor performance. izontal displacement modeling of hydropower dam, Neural Comput. Appl. 29 (12)
(5) The proposed SEV-MFOA-SVM model is a novel approach to pre­ (2018) 1495–1506.
[20] X. Li, Z. Wen, H. Su, An approach using random forest intelligent algorithm to
dicting and analyzing dam deformation, and potentially other si­
construct a monitoring model for dam safety, Eng. Comput. (2019) 1–18.
milar structural behavior, however, the measured temperature ef­ [21] F. Kang, J. Li, S. Zhao, et al., Structural health monitoring of concrete dams using
fects on dam body is not considered in this model. Another long-term air temperature for thermal effect simulation, Eng. Struct. 180 (2019)
642–653.
significant issue is that the proposed model is essentially static.
[22] F. Salazar, M.Á. Toledo, J.M. González, et al., Early detection of anomalies in dam
Further studies are needed to better the accuracy and reliability of performance: A methodology based on boosted regression trees, Struct. Control
dam deformation prediction with respect to the SEV-MFOA-SVM Health Monitor. 24 (11) (2017) e2012.
model. [23] C. Lin, T. Li, S. Chen, et al., Gaussian process regression-based forecasting model of
dam deformation, Neural Comput. Appl. (2019) 1–16.
[24] D.T. Bui, N.D. Hoang, H. Nguyen, et al., Spatial prediction of shallow landslide
Declaration of Competing Interest using bat algorithm optimized machine learning approach: A case study in Lang Son
Province, Vietnam, Adv. Eng. Informat. 42 (2019) 100978.
[25] N.D. Hoang, Image processing based automatic recognition of asphalt pavement
The authors declare that they have no known competing financial patch using a metaheuristic optimized machine learning approach, Adv. Eng. Inf. 40
interests or personal relationships that could have appeared to influ­ (2019) 110–120.
[26] M. Golparvar-Fard, A. Heydarian, J.C. Niebles, Vision-based action recognition of
ence the work reported in this paper. earthmoving equipment using spatio-temporal features and support vector machine
classifiers, Adv. Eng. Inf. 27 (4) (2013) 652–663.
Acknowledgements [27] H. Zhang, S.F. Xu, Multi-scale dam deformation prediction based on empirical mode
decomposition and genetic algorithm for support vector machines (GA-SVM),
Chinese J. Rock Mechanics Eng. 30 (S2) (2011) 3681–3688.
This research was jointly funded by the National Key Research and [28] Wan Z, Dong H, Liu B. On choice of hyper-parameters of support vector machines
Development Program (Grant No. 2018YFC0406905) and the National for time series regression and prediction with orthogonal design. Rock and Soil
Mechanics, 2010, 31(2): 503-508+515.
Natural Science Foundation of China (Grant No. 51879185). Also, the
[29] Y. Li, L. She, L. Wen, et al., Sensitivity analysis of drilling parameters in rock rotary
authors gratefully acknowledge the valuable suggestions of the editor drilling process based on orthogonal test method, Eng. Geol. 105576 (2020).
and the three anonymous reviewers. [30] W.T. Pan, A new fruit fly optimization algorithm: Taking the financial distress
model as an example, Knowl.-Based Syst. 26 (2012) 69–74.
[31] H. Li, S. Guo, H. Zhao, et al., Annual electric load forecasting by a least squares
References support vector machine with a fruit fly optimization algorithm, Energies 5 (11)
(2012) 4430–4445.
[1] J. Jia, A technical review of hydro-project development in China, Engineering 2 (3) [32] S.M. Lin, Analysis of service satisfaction in web auction logistics service using a
(2016) 302–312. combination of fruit fly optimization algorithm and general regression neural

15
Q. Ren, et al. Advanced Engineering Informatics 46 (2020) 101154

network, Neural Comput. Appl. 22 (3–4) (2013) 783–791. (11) (2018) 1328–1338.
[33] B. Dai, C. Gu, E. Zhao, et al., Statistical model optimized random forest regression [46] S. Lu, Y. Wang, Y. Wu, Novel high-precision simulation technology for high-dy­
model for concrete dam deformation monitoring, Struct. Control Health Monitor. namics signal simulators based on piecewise Hermite cubic interpolation, IEEE
25 (6) (2018) e2170. Trans. Aerosp. Electron. Syst. 54 (5) (2018) 2304–2317.
[34] M. Li, Q. Ren, R. Kong, et al., Dynamic modeling and prediction analysis of dam [47] H. Khosroshahi, S.M.M. Husseini, M.R. Marjani, The bullwhip effect in a 3-stage
deformation under multidimensional complex relevance, J. Hydraul. Eng. 50 (6) supply chain considering multiple retailers using a moving average method for
(2019) 687–698. demand forecasting, Appl. Math. Model. 40 (21–22) (2016) 8934–8951.
[35] J. Zhang, J. Wang, L. Chai, Factors influencing hysteresis characteristics of concrete [48] J. Ye, Improved cosine similarity measures of simplified neutrosophic sets for
dam deformation, Water Sci. Eng. 10 (2) (2017) 166–174. medical diagnoses, Artif. Intell. Med. 63 (3) (2015) 171–179.
[36] S. Wang, Y. Xu, C. Gu, et al., Hysteretic effect considered monitoring model for [49] P. Castagliola, A. Achouri, H. Taleb, et al., Monitoring the coefficient of variation
interpreting abnormal deformation behavior of arch dams: A case study, Struct. using a variable sample size control chart, Int. J. Adv. Manuf. Technol. 80 (9–12)
Control Health Monitor. (2019) e2417. (2015) 1561–1576.
[37] B. Wei, M. Gu, H. Li, et al., Modeling method for predicting seepage of RCC dams [50] D.N. Reshef, Y.A. Reshef, H.K. Finucane, et al., Detecting novel associations in large
considering time-varying and lag effect, Struct. Control Health Monitor. 25 (2) data sets, Science 334 (6062) (2011) 1518–1524.
(2018) e2081. [51] Y. Zhang, W. Zhang, Y. Xie, Improved heuristic equivalent search algorithm based
[38] Song Z, Li J. Research on dam displacement forecasting model based on least on maximal information coefficient for Bayesian network structure learning,
squares support vector machine. Water Resources and Power, 2006, (6): 49- Neurocomputing 117 (2013) 186–195.
52+115-116. [52] R. Kromanis, P. Kripakaran, Support vector regression for anomaly detection from
[39] M.M.R. Tabari, H.R.Z. Sanayei, Prediction of the intermediate block displacement measurement histories, Adv. Eng. Inf. 27 (4) (2013) 486–495.
of the dam crest using artificial neural network and support vector regression [53] A. Dimitrov, M. Golparvar-Fard, Vision-based material recognition for automated
models, Soft. Comput. 23 (19) (2019) 9629–9645. monitoring of construction progress and generating building information modeling
[40] Wang X, Fan Q, Xu C, et al. Dam deformation prediction based on wavelet transform from unordered site image collections, Adv. Eng. Inf. 28 (1) (2014) 37–49.
and support vector machine. Geomatics and Information Science of Wuhan [54] J.S. Chou, A.D. Pham, Smart artificial firefly colony algorithm-based support vector
University, 2008, (5): 469-471+507. regression for enhanced forecasting in civil engineering, Comput.-Aided Civ.
[41] H. Su, Z. Wen, X. Sun, et al., Time-varying identification model for dam behavior Infrastruct. Eng. 30 (9) (2015) 715–732.
considering structural reinforcement, Struct. Saf. 57 (2015) 1–7. [55] M. Liu, Z. Cao, J. Zhang, et al., Short-term wind speed forecasting based on the
[42] H. Su, Z. Wen, X. Sun, et al., Rough set-support vector machine-based real-time Jaya-SVM model, Int. J. Electr. Power Energy Syst. 121 (2020) 106056.
monitoring model of safety status during dangerous dam reinforcement, Int. J. [56] G. Cao, L. Wu, Support vector regression with fruit fly optimization algorithm for
Damage Mech. 26 (4) (2017) 501–522. seasonal electricity consumption forecasting, Energy 115 (2016) 734–745.
[43] H. Su, Z. Wen, Z. Chen, et al., Dam safety prediction model considering chaotic [57] L. Shen, H. Chen, Z. Yu, et al., Evolving support vector machines using fruit fly
characteristics in prototype monitoring data series, Struct. Health Monitor. 15 (6) optimization for medical data classification, Knowl.-Based Syst. 96 (2016) 61–75.
(2016) 639–649. [58] D.K. Bui, T. Nguyen, J.S. Chou, et al., A modified firefly algorithm-artificial neural
[44] B. Wei, L. Chen, H. Li, et al., Optimized prediction model for concrete dam dis­ network expert system for predicting compressive and tensile strength of high-
placement based on signal residual amendment, Appl. Math. Model. 78 (2019) performance concrete, Constr. Build. Mater. 180 (2018) 320–333.
20–36. [59] M. Li, Y. Shen, Q. Ren, et al., A new distributed time series evolution prediction
[45] M. Li, Q. Ren, Y. Shen, Prediction model for interactive time series evolution and its model for dam deformation based on constituent elements, Adv. Eng. Inf. 39 (2019)
verification of dam deformation under Bayesian framework, J. Hydraul. Eng. 49 41–52.

16

You might also like