Received 29 January 2025, accepted 10 February 2025, date of publication 14 February 2025, date of current version 4 March 2025.

Digital Object Identifier 10.1109/ACCESS.2025.3542355

Power Transformer Vibration Analysis Model

Based on Ensemble Learning Algorithm
BAIDI SHI 1 , YONGFENG JIANG2 , WEI XIAO3 , JINGYU SHANG1 , MENG LI3 , ZIXING LI3 ,
AND XINFU CHEN3
1 Collegeof Mechanical and Electrical Engineering, Hohai University, Changzhou, Jiangsu 213001, China
2 Collegeof Materials Science and Engineering, Hohai University, Changzhou, Jiangsu 213001, China
3 Changzhou Xidian Transformer Company Ltd., Changzhou, Jiangsu 213017, China

Corresponding authors: Yongfeng Jiang ([email protected]) and Baidi Shi ([email protected])

This work was supported in part by the National Science and Technology Major Project under Grant 2024ZD0803200; in part by the
Postgraduate Research and Practice Innovation Program of Jiangsu Province under Grant KYCX24_0896; in part by the Fundamental
Research Funds for the Central Universities under Grant B240205008; in part by the National Natural Science Foundation of China under
Grant 51879089; and in part by the Cooperative Innovational Center for Coastal Development and Protection, for the first group, 2011 Plan
of China’s Jiangsu Province under Grant 201356.

ABSTRACT The inflow of DC magnetic flux into the core of power transformers under DC bias conditions
leads to a significant increase in operational losses, noise levels, and vibration amplitudes, posing substantial
threats to the stability and reliability of the power system. To comprehensively examine the effects of
DC bias on the vibration characteristics of power transformers, a fully coupled 3D finite element model
of the transformer’s electromagnetic-structural physical fields was developed. This model was validated
through experimental comparison, confirming its accuracy in representing physical behaviors. In the second
phase, a parametric finite element model, combined with the Sobol method, was utilized to design a set of
samples for sensitivity analysis, enabling a quantitative study of the interactions among key characteristic
parameters. Finally, a prediction model based on ensemble learning was proposed, using the Stacking method
to integrate decision tree regression, support vector machine, and extreme learning machine algorithms.
Three mainstream heuristic algorithms were applied for hyperparameter optimization, with the genetic
algorithm (GA) proving to best meet the requirements, limiting the maximum prediction error to 1.2 dB.
The model was validated on two in-service transformers operating under extreme conditions—DC bias and
significant load fluctuations, the maximum error is less than 3 dB, demonstrating certain generalizability.

INDEX TERMS Power transformer, multi-physical field coupling, global sensitivity analysis, dc bias,
ensemble learning, machine learning, hyperparameter optimization.

I. INTRODUCTION storms [4], stray currents [5], and monopole operation [6]
Based on the principle of electromagnetic induction [1], [2], in Ultra-High Voltage Direct Current (UHVDC) transmission
power transformers are electrical devices that connect circuits systems. When the DC component flows through the wind-
with different voltage levels in power systems, serving as ing, it distorts the excitation current, generates harmonics,
indispensable components in AC power grids. During routine significantly increases transformer losses, vibration ampli-
testing [3], the primary side of the transformer is typically tudes, and operational noise, and threatens the stability and
excited with a standard power-frequency sine wave. How- reliability of the transformer.
ever, in actual operation, the neutral end of the transformer The vibration sources of power transformers are complex,
can be affected by DC components due to solar magnetic with their vibration frequencies being mixed and coupled,
primarily originating from the core, winding, and cooling
The associate editor coordinating the review of this manuscript and equipment. The vibration process is nonlinear, influenced
approving it for publication was Tianhua Xu . by the coupling of electromagnetic and structural [7], [8]

2025 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
37812 For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 13, 2025
B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

multi-physical fields, making its analysis intricate. Currently, computational complexity of multi-physical finite element
some researchers have focused on the vibration of trans- analysis, [20] proposed a regression equation for transformer
former windings, often analyzing it using structural dynamics load vibration, using operating capacity and excitation fre-
principles [9]. In these studies, the winding and its structural quency as independent variables. Reference [21] established
components are modeled as a multi-degree-of-freedom mass- an electromagnetic-structural coupled finite element analy-
spring-damping system. Although analytical methods offer sis model for a 220 kV/240 MVA power transformer and
the advantages of simplicity and closed-form solutions, they conducted a detailed analysis of the effects of load and mea-
have limitations. Due to the nodal equivalence of mechanical surement point height on the vibration of the transformer
structures, these methods fail to account for the nonlinear tank. However, most existing studies [15], [16], [17], [18],
characteristics of materials and the unique local features of [19], [20], [21] focus on specific phenomena in transformers
structural components. To address these shortcomings, the and conduct isolated modeling analyses. Currently, no com-
multi-physical field coupling finite element method (FEM) prehensive model exists to systematically and quantitatively
has been applied to analyze winding vibration under electro- analyze the effects of varying capacities, voltage levels, exci-
magnetic and structural coupling [10], [11], [12], [13]. As a tation currents, and DC components on transformer vibration
numerical method, FEM accurately calculates stress distribu- and operating noise.
tion in any winding section by using interpolation functions In recent years, machine learning algorithms have been
and setting appropriate boundary conditions. It can also solve widely used in engineering due to their robustness and nonlin-
transient characteristics and nonlinear relationships by defin- ear fitting ability [22], [23]. While machine learning performs
ing constitutive relations and appropriate time steps. For less effectively than deep learning in big data applications,
example, in [14], an electromagnetic-mechanical coupling it remains valuable in small-sample learning, especially in
model was developed to evaluate winding vibration, current, engineering, where dataset creation is time-consuming and
and magnetic leakage under unbalanced operation. While expensive. In [24], Decision Tree Regression (DTR) com-
winding vibration contributes to transformer internal vibra- bined with an adaptive enhancement algorithm was used for
tion, it can be effectively suppressed by using high-damping optimal control and state prediction of an eco-water symbio-
plates and adjusting the winding preload, making it a sec- sis system. Similarly, in [25], Support Vector Machine (SVM)
ondary source of vibration and noise. Additionally, ensuring was applied for fault diagnosis in lithium batteries, show-
model accuracy requires significant computational resources ing higher accuracy than other machine learning algorithms.
and time, which limits the feasibility of FEM for real-time Moreover, machine learning models offer great potential for
evaluation or use during the design phase. integration with other algorithms due to their low computa-
The vibration of the transformer core is primarily caused tional complexity. In [26], a model using Extreme Learning
by the inherent magnetostriction properties of ferromag- Machine (ELM) and fuzzy control was developed to evaluate
netic materials [15] and the electromagnetic forces between solar panel operating states, improving the stability of new
core layers [16]. Under normal sinusoidal excitation, the energy systems. However, as noted by the ‘‘no free lunch’’
vibration frequency is an even multiple of the AC current theorem [27], [28], no single algorithm excels across all
frequency, exhibiting periodic regularity. However, under DC problems. Literature [27] suggests that adopting an ensemble
demagnetization, direct current flux intrudes into the core, learning strategy, which integrates multiple heterogeneous
leading to half-wave saturation, distortion of the excitation models, can overcome the limitations of a single algorithm
current, and a rapid increase in core vibration. This is the and improve overall model performance. Ensemble learning
primary cause of abnormal vibration and noise in trans- methods [29], [30], [31] have been successfully applied in
formers. In [17], experiments revealed the precise magnetic wind power prediction, online load forecasting, and resi-
strain curve of silicon steel sheets, showing that noise and dential net load interval prediction, achieving better results
vibration significantly increase under non-sinusoidal exci- than single models. The vibration characteristics of power
tation compared to sinusoidal conditions. Since the core transformers are determined by internal design parameters
consists of multi-scale silicon steel sheets, abnormal noise and operating conditions, making the evaluation process non-
and vibration are further intensified under non-sinusoidal linear and complex. However, to the best of our knowledge,
excitation, particularly with DC bias, posing serious threats few studies have applied ensemble learning frameworks to
to the safety and stability of the power system. In [18], the transformer design, optimization, and diagnostics.
vibration half-wave energy method was introduced as an In light of the limitations of existing research, this paper
evaluation index to assess DC bias in transformers, validated proposes the following studies: (a) Based on the Sobol sample
using FEM combined with experimental results. Similarly, design principle, a set of power transformer vibration sam-
[19] employed a field coupling method to determine coil ples is generated using a parametric finite element model,
node currents, effectively reducing the computational com- and sensitivity analysis is performed to explore the relation-
plexity of three-dimensional transient electromagnetic field ship between various parameters and transformer vibration;
analysis and obtaining the magnetic density distribution of (b) An ensemble learning model integrating Support Vector
the core under DC bias conditions. To address the high Machines (SVM), Decision Tree Regression (DTR), Linear

VOLUME 13, 2025 37813

 B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

Regression (LR), and Extreme Learning Machine (ELM) The above formula is the electromagnetic-mechanical cou-
is developed to evaluate transformer vibration using the pling governing equations, by introducing the FEA method,
obtained dataset; (c) Hyperparameter optimization is con- the magnetic field intensity, the electromagnetic force and
ducted for the model, and engineering validation under other field values of each node can be obtained. Figure 1
special operational conditions is performed to further assess shows the.
the model’s robustness and practical applicability.
B. MODEL BUILDING
II. EXPERIMENTAL DESIGN Machine learning models are all data-driven, meaning that
A. MUTI-PHYSICS COUPLING THEORY the precision and completeness of the sample data will
Considering many finite element calculations are carried out directly determine the overall performance of the model.
to build the FEA dataset, the magnetic vector potential A The main source of vibration in power transformers is the
is utilized to simplify the calculational complexity. Based core, which is transmitted to the oil tank through struc-
on Maxwell’s equations and field coupling principle, the tural components such as clamps and feet. This process is
constitutive relationship between the external current density an electromagnetic-structural coupling process. Combining
J e and the magnetic flux density B is, existing research results [15], [16], [17], [18], [19], [20],
∂A   the design variables selected in this paper are related to the
σ + ∇ × µ(H)−1 ∇ × A = J e (1) transformer core and electrical parameters. Their specific
∂t
types and descriptions are listed in Table 1.
where σ and H mean, respectively, the electrical conductivity
of the material and the magnetic field strength vector. µ(·)−1 TABLE 1. Design parameters.
is the magnetic permeability, which is introduced to describe
the nonlinear magnetization process.
The vibration transient governing equation of the trans-
former can be characterized by the structural dynamic
differential equation, which is,
d 2 S(t) dS(t)
M 2
+C + KS(t) = F(t) (2)
dt dt
where M is the mass coefficient matrix; C is the damping
coefficient matrix; K is the stiffness coefficient matrix; S is
the displacement of the node, and its corresponding first and
second derivatives are its velocity and acceleration. F(t) is
the external force matrix of the node, which is mainly caused
by Maxwell’s electromagnetic force and magnetostriction
In Table 1, variable BH is a set of discrete coordinate
force in the electromagnetic coupling field. Among them,
points, which is the sampling point of the magnetization pro-
Maxwell’s electromagnetic force applied in different direc-
cess of the silicon steel sheet measured by the manufacturer.
tions can be expressed as.
 Taking a certain brand of silicon steel sheet as an example, the
 fx = Jy Bz − Jz By
 general form of the variable can be seen in Figure 1 below.
fy = Jz Bx − Jx Bz (3) The variable T describes the core type, and its values can
 be 1, 2, and 3 corresponding to single-phase, three-phase,
fz = Jx By − Jy Bx

and three-phase five-pillar, in turn. The model in this study
where Bx , By , and Bz are the magnetic induction intensity in x, primarily targets large-capacity power transformers, which
y, and z direction;, Jx , Jy and Jz are the three direction current are typically oil-immersed. Although dry-type transformers
densities corresponding to x, y, and z. (e.g., 110 kV/50,000 kVA) exist, they are not widely adopted.
For the magnetostrictive properties of silicon steel, the Regarding the core design, both core-type and shell-type
deformation λ of silicon steel sheet under different magneti- cores can be suitable for various voltage levels and capacities.
zation intensities can be characterized by a secondary domain However, in the power system analyzed in this study, the
transfer model, which can be expressed as, core-type design is preferred due to its widespread use in
   2 large-capacity transformers.
3 1 3 Mi 1
λi = λs αi2 − = λs [ − ] (4) It can be known that each design parameter involves var-
2 3 2 Ms 3 ious data types and has significant differences. Traditional
where i is the orientation identifier; λs is the magnetostriction test methods such as orthogonal tests and center point tests
coefficient of the silicon steel sheet under saturation state, and require that each independent variable has a clear value range
its unit is ppm; Mi and Ms mean, respectively, the magnetiza- and the same data type, which cannot meet the actual needs
tion in the selected direction and the saturation magnetization of this project. The test design scheme in this project is
of the material. generated using the Sobol sequence [27]. This method is

37814 VOLUME 13, 2025

B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

based on the principle of discretized Monte Carlo sampling of the analysis. Most conventional transformers exhibit
and is oriented towards maximizing the variance. It performs near-symmetric geometric structures and electromag-
parameter combinations for discrete or continuous indepen- netic distributions, particularly on the high-voltage side.
dent variables in the design process, making the parameter Additionally, considering the need for large-scale model-
combination points evenly distributed within the domain, and ing to establish a comprehensive sample set in subsequent
solves the design problems of complex engineering schemes. studies, it is essential to balance computational com-
Based on the aforementioned Sobol sequence generation plexity, accuracy, and practical feasibility. This approach
method, a total of 1,576 experimental combinations were ensures the model remains efficient and scalable for
generated. The distribution of the experimental schemes is extended analyses.
well-balanced across the design space, effectively addressing (c) The materials for the core, windings, and tank were set
the issue of sample completeness. Furthermore, to ensure the as silicon steel (with permeability adaptively interpolated
accuracy of the dataset, all generated experimental combina- based on discrete BH points), copper, and Q235 structural
tions were validated and solved using finite element software. steel, respectively.
The overall solution process is illustrated in Figure 1. (d) The surfaces of structural components such as the tank
were assigned impedance boundary conditions to reduce
the complexity of finite element computations.
(e) A tetrahedral adaptive meshing strategy was employed,
with a mesh size of 80 mm for the air domain and 40 mm
for other components.
(f) Excitation was applied using a field-circuit coupling
method, with the excitation type set as a sinusoidal volt-
age source. A DC bias component was added to the
excitation source to simulate DC offsets in the excitation.
(g) To ensure the validity of analysis within the [0,1000]
Hz frequency range, the Nyquist sampling theorem was
adhered to. The simulation duration was set to 0.2 s, with
a time step of 1e-4s.
Following the above configurations and computations,
acceleration signals at each node within the finite element
domain were obtained for the specified time interval. Addi-
tionally, through quadratic interpolation, these results were
extended from the quadrilateral vertices to all coordinate
points within the solution domain. Finally, the output features
for each sample were selected as the signals at nine measure-
ment points located on the surface of the tank, directly facing
either the center column or side columns. A specific example
FIGURE 1. Solution process. of these measurements is provided in Figure 2.

The finite element sample generation process is depicted

in the figure. Considering the complexity of solving tran-
sient coupled electromagnetic fields, the following settings
and simplifications were applied to the parameterized finite
element model:
(a) Bolt connections within structural components were
omitted due to the complexity of accurately modeling
preloading forces, contact behaviors, and interactions
between components like the core, wooden parts, and
iron structures. This approach reduces computational
demand and aligns with references [15], [17], [18],
[21], which similarly exclude bolt connections. While
bolt connections may influence vibration behavior, their
omission focuses the model on primary vibration factors.
This limitation is acknowledged in the discussion section, FIGURE 2. Point measurement.
with future research proposed to address it.
(b) A symmetry-based half-model was adopted to reduce It is worth noting that both the sampled signals and the
computational complexity while retaining the accuracy finite element analysis (FEA) results are time series data.

VOLUME 13, 2025 37815

 B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

To validate the accuracy of the FEA results, the time-domain maximum sampling frequency of 26.7 kHz. According to the
vibration acceleration signals at the measurement points were inspection results from the certified testing center, the max-
extracted and subjected to Fourier transform. Subsequently, imum measurement error within the [0, 2000] Hz frequency
the acceleration amplitudes obtained from the Fourier trans- range is less than 1.7%, demonstrating its suitability for accu-
form of each measurement point were weighted according to rately measuring transformer vibration signals. Additionally,
Equation (5) and used as the quantitative evaluation metric to minimize the impact of high-frequency noise caused by
for vibration acceleration. environmental disturbances, a low-pass filter was applied
 Va1 Va2 Van to retain frequency components below 2000 Hz, ensuring
Vi 10 + 10 10 . . . . . . + 10 10 )
asum = 10log10 (10 that the data reflects the core vibration characteristics of the
i ai (5)
 Vidx = 20lg transformer. Data transmission in this study was implemented
a0 using the Modbus-TCP protocol, which avoids the need
where ai represents the effective acceleration at the i-th fre- for expensive multi-channel data acquisition devices. Time-
quency, while a0 denotes the reference acceleration value, domain synchronization of signals from nine measurement
set to 10−6 m/s2 in this study. Vidxi corresponds to the ref-
points was performed, achieving a synchronization precision
erence acceleration value at the i-th frequency for the idx of ±1 ms. This ensures that signals from all measurement
measurement point. Following the above processing steps, the points are aligned in time, enabling accurate analysis of vibra-
output of any sample is represented as a 9-dimensional vector, tion characteristics. The Modbus-TCP protocol also offers
corresponding to the nine measurement points selected in advantages such as high reliability, ease of integration with
Figure 2. Since these measurement points are located on the existing network infrastructure, and the ability to efficiently
surface of the transformer tank, the measurement process is handle multi-node communication. This setup allows the host
straightforward and enables direct comparison with actual computer to directly read the characteristic signals of each
signals, thereby validating the effectiveness of the sample set. numbered sensor within the same network segment, signif-
icantly simplifying the data acquisition process. Figure 4
III. NUMERICAL ANALYSIS illustrates the comparison of measured and predicted vibra-
A. DATASET VALIDITY ANALYSIS tion signals at each measurement point of the transformer
In Chapter 1, a finite element sample set for vibration sig- under three different operating conditions: (a) stable 65%
nals of power transformers was established. In this chapter, load with 4A dc bias, (b) no-load condition, and (c) 1.1 over-
the validity and accuracy of this sample set are verified to load condition. Each subfigure provides insights into the
demonstrate its capability to effectively describe the vibration alignment between the measured data and the finite ele-
characteristics of power transformers. The test case selected ment model predictions, demonstrating the applicability of
is a 500 kV/334 MVA single-phase auto-transformer from the sample set across various scenarios:(a) As shown in
a specific manufacturer, The arrangement of measurement Figure 4(a), under a stable 65% load condition with dc bias,
points is consistent with that shown in Figure 3. the maximum error is less than 5%, demonstrating the finite
element model’s strong predictive capability under general dc
bias conditions and (b) Under special operating conditions,
such as no-load and 1.1 overload scenarios, the model’s pre-
diction errors tend to increase but remain within a maximum
error of 8%. This suggests that even under special operating
conditions, the model provides reliable and referenceable
predictions.

B. SENSITIVITY ANALYSIS
The Sobol method [32] is a global sensitivity analysis
approach based on variance decomposition. In this method,
for an input vector Xi with dimensionality d, the output fi (Xi )
can be decomposed into the following components:
• First-order parameter terms, fi (Xi ), which quantify the
contribution of individual input variables.
• Interaction terms, fij (Xi , Xj ), which represent the inter-
actions between pairs of variables.
• Higher-order interaction terms, f1,...,k (X1 , X2 ,. . . , Xj ,),
FIGURE 3. Data processing flow. where j=3,. . . , d.
This hierarchical decomposition provides a comprehensive
The acceleration sensor used in this study is the VTall- understanding of how each input variable and their interac-
T163E-FC model from a specific manufacturer, with a tions contribute to the output variance, the output f (X) can be

37816 VOLUME 13, 2025

B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

By combining Equations (2) and (3), the discrete variance

expression can be derived as:
 Xd Xd
 Var (Y ) =

 Vi + Vi,j + . . . + V1,2,...,d
i=1 i<j
 Vi = Var Xi (EX ∼i (Y |Xi ))
Vi,j = Var Xi,j EX ∼ij Y | Xi , Xj − Vi − Vj



(8)
where Var(·) represents the variance calculation function, and
EX ∼i (·) denotes the mathematical expectation of the iii-th
factor. Equation (8) can be used to calculate the variance con-
tributions of variables at different orders, including first-order
terms, higher-order terms, and interaction terms. These repre-
sent the impacts of independent variable changes, interactions
between variables, and higher-order effects on the model,
respectively. For the first-order index of the i-th variable, it is
directly calculated using a variance-based approach and can
be expressed as follows:
Vi
Si = (9)
Var(Y )
where Si represents the contribution of a first-order index to
the output variance. It should be noted that if the variance
decomposition method shown in Equation (8) is used, the
number of indices to evaluate grows exponentially with the
number of input variables. Specifically, for a d-dimensional
input, the total number of sensitivity indices is 2d -1. This
exponential growth in complexity leads to high time and
space costs for the decomposition process, making the anal-
ysis challenging when the system has a high dimensionality.
To address this issue, this study introduces a global evaluation
index to quantitatively assess the overall impact of a specific
input variable on the final output. This includes all interac-
tions between the variable and any other input variables at
any order,
Var X ∼i (EXi (Y |X ∼i ))
STi = 1 − (10)
Var(Y )
FIGURE 4. Comparison under three conditions. where STi represents the global sensitivity of the i-th input
variable to the model output. The Sobol method, based on
rewritten in a variance form, the principle of variance decomposition, can calculate the
Xd Xd impact of terms of any order on the final output, offering high
fi (Xi ) + fij Xi , Xj + · · ·


f (X) = f0 + flexibility and suitability for engineering applications. Using
i=1 i<j
+ f1,2,...,d (X1 , X2 , . . . ,Xd ) (6) the Sobol method described above, the global sensitivities of
the design parameters listed in Table 1 to the transformer
In the equation, f0 represents the constant term, while vibration were computed. In the original sample set, the
fi,jis a function of variables Xi and Xj , describing the effect output variables correspond to the nine measurement points
of second-order interactions between factors on the output shown in Figure 2. To reduce computational complexity,
variance. All parameter terms are pairwise orthogonal, and the average values of three measurement points at the same
the decomposition satisfies the following condition: height were used, reducing the final output dimensionality to
Z 1 three. Subsequently, three separate global sensitivity analyses
fi1 i2 ...is Xi1 , Xi2 , . . . ,Xis dXk = 0


were conducted. The results of these analyses are summarized
0
in Figure 5.
f0 = E(Y )
Figure 5 summarizes the results of the three sensitivity
fi (Xi ) = E (Y | Xi ) − f0 analyses. It can be observed that at all three heights, the DC
fij Xi , Xj = E Y | Xi , Xj − f0 − fi − fj



(7) bias component and the core design magnetic flux density are

VOLUME 13, 2025 37817

 B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

where c1 and c2 represent the fixed output values for the

partitioned regions R1 and R2 , respectively. Based on the least
squares method, to achieve minimum fitting variance, c1 and
c2 must satisfy the following conditions,
1 X

 c1 =
 yi
N1 iϵR1
(12)
1 X
 c2 =
 yi
N2 iϵR2

where N1 and N2 represent the number of samples in regions

R1 and R2 , respectively. For the generated subregions R1
and R2 , the process described by Equation (11)-(12) will be
recursively applied until the iteration criteria are met. Finally,
M partitioned subresions R1 , R2 ,. . . , RM are obtained, and the
output can be expressed as,
FIGURE 5. Sensitivity analysis.
XM
DTR (x) = cm I (x ∈ Rm ) (13)
m=1
the primary factors influencing the vibration amplitude of the
power transformer. Moreover, the global sensitivities exhibit where cm denotes the fixed output value of subregion Rm , and
certain variations with increasing measurement height. As the I (·) means the indicator function that equals 1 if x belongs to
height of the measurement points increases, the influence of the region Rm , and 0 otherwise.
transformer dimensional parameters on the vibration ampli- The Decision Tree Regression (DTR) model is a widely
tude becomes more significant. used algorithm in regression tasks due to its simplicity and
interpretability. However, the performance of DTR is highly
IV. ENSEMBLE LEARNING MODEL sensitive to its hyperparameters, making careful adjustment
Ensemble learning algorithms train and integrate diverse essential for achieving optimal results. Key hyperparameters,
base learners, effectively leveraging their individual strengths such as maximum tree depth (MTD), minimum samples split
while mitigating their inherent limitations. This approach (MSS), and minimum samples leaf (MSL), directly con-
aims to achieve improved overall performance. In this chap- trol the model’s complexity. Improper settings may lead to
ter, we first introduce principles of the basic learners and meta overfitting or underfitting, significantly affecting prediction
learner and then describe the integrated learning strategy of accuracy. Furthermore, the choice of the splitting criterion
Stacking. (e.g., mean squared error or mean absolute error) impacts the
quality of splits and the model’s ability to capture underlying
A. DECISION TREE REGRESSION data patterns. Thus, fine-tuning these hyperparameters not
Decision Tree Regression (DTR) is a classic machine learn- only balances the trade-off between model complexity and
ing regression model. During the training process, it partitions generalization but also ensures robust and reliable predictions
the feature space based on specific principles, achieving high in diverse datasets.
predictive accuracy with relatively low time and space com-
plexity. In this study, DTR model employs the mean squared B. EXTREME LEARNING MACHINE
error (MSE) as the splitting criterion, which minimizes the
Extreme Learning Machine (ELM) is a two-layer feedfor-
variance of the target variable within each split, ensuring
ward neural network with a fully connected structure similar
better predictions for continuous output variables. Addition-
to that of deep learning networks. However, the input weights
ally, overfitting can be mitigated by limiting the number of
and bias coefficients of the hidden layer in ELM are randomly
node splits and by carefully selecting the training sample
initialized and remain unchanged during training. Unlike tra-
sampling strategy. In this algorithm, a recursive approach is
ditional neural networks that rely on iterative gradient-based
used to heuristically partition the sample space. Each split is
optimization, ELM updates the weights of the output layer
guided by the goal of minimizing the squared error, selecting
in a single step using the least squares method, resulting in
the optimal split point in the sample space. Given the i-th
extremely high training efficiency. Due to the random initial-
feature variable xi from the training set, if its split threshold
ization of input layer coefficients, ELM demonstrates strong
is sss, the sample space can be divided into two subregions
generalization capabilities across various problem domains.
R1 (j, s) = {X | xi ≤ s} and R2 (j, s) = {X | xi > s}. For the
Moreover, the one-step training of output layer weights via
regression model, the optimization objective is to minimize
the least squares method ensures both training efficiency and
the fitting variance, expressed as:
robust nonlinear fitting capabilities. The key computational
step involves determining the weights from the hidden layer
X X
min [min (yi − c1 )2 + min (xi − c2 )2 ]
j,s c1 xi ∈R1 c2 X i ∈R2 to the output layer. The output of the i-th node tˆi in the output
(11) layer can be expressed as a linear combination, given the

37818 VOLUME 13, 2025

B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

known output weights β and hidden layer outputs Hi , the model’s performance. Thus, employing heuristic opti-
( XL mization methods not only enhances the predictive accuracy
tˆi = β l hil = β T hi and robustness of SVM but also reduces the computational
l=1 (14)
H i = g(xW i + bi ) cost associated with exhaustive search techniques.
where W and b represent the input weights and bias bias
D. LINEAR REGRESSION
matrix of the hidden layer, respectively. Commonly, these
The first three chapters introduce three machine learning
parameters are initialized using a Gaussian distribution and
models, which will serve as the first-level learners in the
remain unchanged during training stage. While this design
ensemble learning framework. Considering that these models
enables rapid training and avoids iterative optimization, the
exhibit strong nonlinear data fitting capabilities, it is cru-
random initialization may lead to suboptimal performance,
cial to mitigate potential overfitting. To achieve this, the
as it introduces variability and lacks problem-specific adapt-
second-level learning algorithm re-weights the outputs of
ability. The activation function g(·) is chosen as the Sigmoid
the first-level models, ensuring better overall generaliza-
function. To minimize the fitting variance of the training
tion. In this study, a multivariate linear regression model
set based on the least squares method, β must satisfy the
(Linear Regression, LR) is employed as the second-level
following condition:
learner. To further enhance robustness and prevent overfit-
−1
β = (H T H) HT T (15) ting, regularization terms are incorporated into the regression
model. The optimization objective of this second-level learn-
where T represents the true labels of the training set sam- ing model is formulated as follows:
ples. The Extreme Learning Machine (ELM) is efficient Xn Xm
feedforward neural network, characterized by its randomly obj (W ) = argmin (W xi − yi ) + λ W j (17)
i=1 j=1
initialized hidden layer weights and biases that remain fixed
during training in traditional form. It is of great significant to where W represents the weight matrix assigned to first-level
optimize the initial values of the W and b, in order to enhance models. n and m mean, respectively, the total number of
robustness and convergence of the model. samples and the number of the first layer learner. λ is the regu-
larization coefficient controlling the trade-off between model
C. SUPPORT VECTOR MACHINES complexity and fitting error. This formulation ensures that
The Support Vector Machine (SVM) algorithm uses kernel the second-level learner achieves a balanced aggregation of
functions to map ‘‘low-dimensional non-separable’’ data in the first-level models, thereby improving the generalization
the original space to a higher-dimensional space, thereby capability of the ensemble. The choice of linear regression as
increasing the degrees of freedom and ensuring that the prob- the second-level learner offers enhanced interpretability by
lem becomes linearly separable in the transformed space. directly assigning weights to the first-level models, indicat-
Given a sample set, the regression objective and its constraints ing their relative importance in the ensemble, and ensures
for SVM can be expressed as: simplicity and efficiency by avoiding the added complexity
and computational burden of nested machine learning or
svr (x) = w∗ ϕ (x) + b


 deep learning models, thereby maintaining a transparent and
 2
 w∗ , b∗ = arg mim ∥w∥ + C
 Xm



ξi practical framework for real-world engineering applications.
 w,b,ξi 2 i=1 (16)
(
yi w xi + b ≥ 1 − ξi (∀i = 1, . . . , m)
T E. STACKING STRATEGY



 s.t.

The Stacking method is a general ensemble learning strat-

ξi ≥ 0(∀i = 1, . . . , m)

egy. In this method, base learners are first trained using the
where m represents the number of training samples, while original dataset, and their outputs are used as labels to create
w and b denote the normal vector and bias term of the a new sample set. Subsequently, the second-level learner is
linear plane, respectively, which can be computed based on trained using this new sample set, and its predictions serve
the dual principle and convex optimization theory [23]. C as the final output of the model. To enhance diversity, the
is the regularization coefficient, and ξ is slack variables. Stacking method typically integrates models with different
The choice of the kernel function for the spatial mapping characteristics, achieving generalization through heteroge-
significantly affects the performance of the model. ϕ is the neous ensemble learning. However, it is important to note that
Gaussian kernel function (RBF kernel). However, the per- the use of multi-level learner integration increases the risk
formance of SVM is heavily influenced by the choice of of overfitting. To mitigate this risk, a 5-fold cross-validation
hyperparameters, such as the C, specific kernel parameters approach is employed for sample partitioning in this study.
(e.g., γ for RBF kernels), and the precision threshold ξ in Specifically, the dataset is divided into five equally sized
regression tasks. These hyperparameters control the trade-off folds. In each iteration, four folds are used as the training set
between model complexity, training error, and the ability to to train the base learners, and the remaining fold serves as
generalize to unseen data. Improper hyperparameter settings the validation set to evaluate their performance. This process
can lead to overfitting or underfitting, significantly degrading is repeated five times, with each fold used as the validation

VOLUME 13, 2025 37819

 B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

set once. The outputs of the base learners on the validation is sequentially used as the validation set while the remain-
set are recorded and combined to create the new dataset ing folds serve as the training set. The stacking ensemble
for the second-level learner. The second-level learner is then comprises three first-level learners: Support Vector Machine
trained on this combined dataset to optimize the model’s final (SVM), Decision Tree Regression (DTR), and Extreme
performance. Furthermore, root mean square error (RMSE) Learning Machine (ELM). The predictions from these learn-
is used as the evaluation metric during cross-validation. The ers are fed into a second-level learner, implemented using
RMSE is calculated on the validation set for each fold, and the Linear Regression (LR) with regularization. This hierarchical
average RMSE across all folds is used as the objective value structure leverages the complementary strengths of the base
to guide hyperparameter optimization. This approach ensures learners and improves generalization by re-weighting their
robust performance evaluation and prevents overfitting to the outputs through the second-level learner.
training data. The pseudocode workflow for the Stacking
ensemble learning method is presented in Table 2. V. MODEL VALIDATION
A. MODEL OPTIMIZATION
TABLE 2. Pseudocode of the stacking. This study utilizes Python and its associated libraries for data
processing, model construction, and result analysis. Specifi-
cally, Scikit-learn was employed to implement base learners
such as Support Vector Machine (SVM) and Decision Tree
Regression (DTR), as well as to develop the Stacking Regres-
sor for ensemble learning. KFold cross-validation was used
to evaluate the generalization performance of the models.
NumPy was applied for matrix operations and random ini-
tialization, playing a critical role in the implementation of
Extreme Learning Machine (ELM). Furthermore, this study
manually implemented the ELM algorithm, including ran-
dom initialization of hidden layer weights, activation function
mapping, and least squares optimization of output weights.
Considering that the final output of the stacking model
represents the weighted acceleration values from nine mea-
surement points, the optimization objective is simplified to
the root mean square error (RMSE) across these nine points.
Table 2 summarizes the workflow of the Stacking ensem- This approach reduces the complexity of hyperparameter
ble method. In this study, the first-level learners include tuning and mitigates conflicts among multiple objectives
three machine learning models: Decision Tree Regression in optimization, ensuring a more streamlined and effective
(DTR), Extreme Learning Machine (ELM), and Support model calibration process. Figure 7 illustrates the violin plot
Vector Machine (SVM). These base learners all possess of the root mean square error (RMSE) for the ensemble learn-
strong nonlinear fitting capabilities. To prevent overfitting, ing model under default hyperparameters. Specifically, the
the second-level learning model employs ridge regression ELM algorithm employs 100 hidden neurons with a random
with a penalty term, which linearly aggregates the outputs of initialization strategy.
the first-level models. The workflow for model generation is
illustrated in the Figure 6.

FIGURE 6. Stacking ensemble framework.

In Figure 6, the workflow illustrates a stacking ensemble

framework incorporating a 5-fold cross-validation strategy.
The dataset is first partitioned into five folds, where each fold FIGURE 7. Violin plot of the RMSE.

37820 VOLUME 13, 2025

B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

Figure 7 presents the violin plot of the ensemble learn- function ‘‘flatten’’ of NumPy [33] was utilized to maintain
ing model’s performance on the test samples across each the input form of individual is an one dimension array.
fold under default hyperparameters. It is noteworthy that, 3. For GA, the crossover and mutation rates were set to
due to the completeness of the cross-validation approach, all 0.7 and 0.3, respectively.
samples were involved in both the training and validation 4. For PSO, a linearly decaying inertia factor was used,
phases, ensuring a comprehensive evaluation of the model’s starting at 0.9 and decreasing to 0.5 by the final iteration.
performance. However, the model’s prediction error exhibits 5. The mutation factor was set to 0.7 in DE.
significant variability, with a maximum deviation reaching The computational experiments in this study were con-
up to 2.4 dB, which fails to meet the accuracy requirements ducted on a system with the following specifications: CPU:
for engineering applications. Therefore, heuristic algorithms AMD R9 7950X, and RAM: DDR5 6400. The model training
will be considered in subsequent work to optimize the and validation process, based on the five-fold cross-validation
model’s hyperparameters and improve prediction precision. principle, took a total of 1874 seconds. Figure 8 illustrates
Key hyperparameters of the stacking model, discussed in the fitness function values and population variance across
Sections III. A through 3.D, are summarized in Table 3, which iterations for each heuristic algorithm. In Figure 8(a), all
provides a comprehensive overview of the selected parameter optimization algorithms demonstrate a significant reduction
values. in the fitting bias of the ensemble learning models during the
initial iterations. A closer inspection of the final 100 iter-
TABLE 3. Key parameters of stacking model. ations reveals that the Genetic Algorithm (GA) achieves
the lowest root mean square error (RMSE), highlighting its
superior regression accuracy on the test dataset. Moreover,
Figure 8(b) indicates that both GA and Particle Swarm Opti-
mization (PSO) exhibit minimal fluctuations in population
variance after 50 iterations, reflecting reduced individual dis-
parities and steady convergence. These findings underscore
the relative effectiveness of GA in optimizing the hyperpa-
rameters of the ensemble learning model in this study. The
proposed approach eliminates the need for the conventional
grid search method, which is characterized by high time and
space complexity, and achieves superior optimization perfor-
mance. By leveraging the heuristic Genetic Algorithm (GA),
the optimization process benefits from its ability to balance
exploration and exploitation within the solution space.
The optimization is conducted using the Scikit-Opt library This enables the algorithm to converge toward optimal or
on the Python 3.11 platform, employing heuristic algorithms near-optimal solutions with reduced computational cost com-
such as Particle Swarm Optimization (PSO), Differential pared to exhaustive search methods. Furthermore, the results
Evolution (DE), and Genetic Algorithm (GA) to optimize demonstrate that GA not only accelerates the hyperparameter
the hyperparameters of the ensemble learning model. The tuning process but also enhances the accuracy and robustness
fitness function is defined as the average RMSE across of the ensemble learning model, underscoring its effective-
five-fold cross-validation, ensuring robust performance eval- ness and adaptability in optimizing machine learning models
uation by reflecting the model’s overall accuracy and for complex datasets. The optimal hyperparameters achieved
reducing overfitting risks. While this approach introduces by the GA are listed in Table 4, and Figure 9 illustrates
additional computational complexity due to repeated model the violin plot of the root mean square error (RMSE) of
training in nested loops, the population size and itera- the given parameters to make a contrast with the original
tion count were restricted to improve efficiency without model.
compromising optimization quality. However, this strategy
introduces additional computational complexity, as fitness TABLE 4. Optimal hyperparameters.
evaluation requires nested loops and repeated training of
machine learning models. To address this, the popula-
tion size and iteration count for heuristic algorithms were
restricted and several techniques have been applied during
operation,
1. The maximum number of iterations and population size
for each algorithm were set to 200 and 50, respectively.
2. The shape weight matrix W and the bias matrix b is Table 4 listed the optimal hyperparameters, in which W
dynamic adjusted according to the melm. Therefore, the and b are numerical matrixes and their specific presentation
input dimension is dynamically changing, the built-in of its value has no practical significance. As illustrated in

VOLUME 13, 2025 37821

 B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

and an absence of significant outliers. This indicates that

the optimized model not only achieves superior prediction
accuracy but also exhibits excellent robustness. The stabil-
ity of the error distribution highlights the reliability of the
optimized ensemble learning framework for practical appli-
cations where consistent performance across diverse datasets
is critical.

B. INTERPRETABILITY ANALYSIS
In previous sections, hyperparameter optimization of the
model has been conducted. To further enhance the inter-
pretability of the proposed ensemble learning framework
and verify the unique contributions of each base learner,
a Pearson correlation coefficient analysis was performed.
This analysis aimed to demonstrate the complementarity and
non-redundancy of the base models—SVM, DTR, LR, and
ELM—within the Stacking framework. By examining the
correlations between the predictions of these models on the
test set, we highlighted their distinct roles and indispensable
contributions to the ensemble. The results confirm that the
heterogeneous integration strategy not only improves predic-
tive accuracy and generalization but also ensures that each
base learner provides unique and irreplaceable value to the
final Stacking model predictions. This approach reinforces
the interpretability of the framework by offering clear insights
into how individual models contribute to the ensemble’s
FIGURE 8. Iterative curve: (a) RMSE (b) Population variance. performance.
n ni=1 xi yi − ni=1 xi ni=1 yi
P P P
pxy = q P q P
n ni=1 xi2 − ( ni=1 xi )2 n ni=1 xi2 − ( ni=1 xi )2
P P

(18)
where n denotes the number of samples in the selected test
set, while x and y represent the predicted values of the
two compared models, respectively. The Pearson correla-
tion coefficient satisfies the commutative property, ensuring
that the resulting correlation coefficient matrix is symmetric.
Figure 10 presents the correlation analysis results, illustrating
the pairwise relationships among the predictive outputs of the
models involved in the ensemble. This symmetric matrix pro-
vides a comprehensive understanding of the complementarity
and independence of each base learner within the Stacking
framework.
As illustrated in Figure 10, the self-correlation for each
FIGURE 9. Violin plot after optimization. model is 1, as expected. Among the first-level learning
models—ELM, SVM, and DTR—their pairwise correla-
tions exceed 0.84, indicating similar predictive capabilities
Figures 7 and 9, the integration of heuristic algorithms for while retaining sufficient diversity. This balance ensures
optimizing the Stacking model leads to a substantial reduc- the effectiveness of the heterogeneous ensemble strat-
tion in test set errors under cross-validation. Specifically, the egy by leveraging complementary strengths of individual
maximum error is reduced to as low as 1.2 dB, showcasing the models. In contrast, the penalized linear regression (LR),
effectiveness of the optimization process in enhancing model although exhibiting weaker fitting capabilities compared to
accuracy. Furthermore, by analyzing the violin plot’s intrinsic the first-level machine learning models, demonstrates signifi-
probability density characteristics, it can be observed that cant differences in correlation coefficients. This characteristic
the numerical deviations are well-centered around the mean, is advantageous in mitigating overfitting by introducing a
with the error distribution demonstrating a narrow spread regularization term to the second-level model. In addition to

37822 VOLUME 13, 2025

B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

were conducted. Using the previously established sam-

ple dataset, the DTR, ELM, SVM, Random Forest (RF),
XGBoost, Stacking-2, Stacking-3, and Fully Connected
Neural Network (FCN) models were trained for perfor-
mance comparison. These models were chosen as horizontal
reference models due to their diverse methodologies, compu-
tational characteristics, and representation of state-of-the-art
machine learning and ensemble strategies, as detailed below:
• DTR was selected as a simple, interpretable baseline
model. Despite its susceptibility to overfitting, it serves
as a fundamental tree-based model, allowing us to
compare how more advanced tree-based methods (e.g.,
RF and XGBoost) improve upon its limitations.
• ELM is a lightweight neural network model known
for its simplicity and efficiency in handling small
FIGURE 10. Correlation coefficient matrix.
sample datasets. Its inclusion provides a contrast
between traditional machine learning models and
neural network-based methods within the ensemble
the correlation analysis of the base model outputs, the inter- framework.
pretability of the framework is further enhanced using a linear • SVM is a widely used machine learning algorithm that
regression model as the second-level learner. This choice leverages kernel functions to handle nonlinear relation-
allows the contribution of each base model to the final out- ships effectively. By including SVM, we evaluate the
put to be directly represented by the regression coefficients. performance of a classical nonlinear model in compari-
Table 5 summarizes the weight distribution of each base son to ensemble-based and deep learning models.
model’s output under different regularization coefficients. • RF is a Bagging-based ensemble method that integrates
multiple decision trees. Its strong generalization capa-
TABLE 5. Weight distribution of basic learners. bility and robust handling of noise make it a relevant
benchmark for comparing the heterogeneous ensemble
strategy employed in the Stacking framework.
• XGBoost represents one of the most advanced gra-
dient boosting algorithms, widely regarded for its
efficiency and effectiveness in both regression and
classification tasks. Including XGBoost ensures the
proposed Stacking framework is evaluated against a
cutting-edge ensemble learning approach, providing a
Table 5 demonstrates the weight distribution of the base rigorous benchmark.
learners (SVM, ELM, and DTR) under varying regulariza- • Stacking-2 employs K-Nearest Neighbors (KNN) as the
tion coefficients (λ) in the linear regression model. The meta-learner. KNN, as a nonparametric model, relies
results indicate that increasing the regularization coefficient on distance metrics rather than explicit assumptions
effectively balances the weight contributions of each model, about data distributions. This configuration helps to
thereby enhancing the robustness and validity of the ensemble explore whether a simpler, nonparametric meta-learner
framework. As λ increases, the dominance of any single can improve ensemble performance.
model diminishes, and the weights tend to become more • Stacking-3 introduces Polynomial Regression as the
evenly distributed. This adjustment ensures that no single meta-learner to account for potential nonlinear inter-
model disproportionately influences the final predictions, actions between the outputs of the base models. This
effectively leveraging the complementary strengths of all configuration evaluates whether including second-order
base learners and preventing overfitting. Furthermore, the terms can enhance the flexibility and accuracy of the
use of a linear regression model as the second-level learner ensemble framework.
enhances the interpretability of the ensemble framework by • FCN serves as a deep learning-based reference model,
allowing the contribution of each base model to be explicitly showcasing the performance of a more complex neu-
represented through its weight. ral network architecture. By comparing FCN with the
proposed Stacking framework, we can highlight the
C. PERFORMANCE COMPARISON trade-offs between computational complexity, model
To highlight the superior performance of the proposed Stack- interpretability, and prediction stability.
ing model compared to other machine learning and deep For all models, hyperparameter optimization was conducted
learning models of the same type, additional experiments using the Genetic Algorithm (GA) to ensure consistency

VOLUME 13, 2025 37823

 B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

and fairness in the comparison. Key parameters such as d) While XGBoost and Stacking-3 delivered competitive
tree depth and number of estimators for RF, kernel type results, their performance was slightly inferior to the pro-
and regularization parameter for SVM, hidden neurons and posed Stacking model, particularly in terms of stability
activation functions for ELM and FCN, and learning rate (higher maximum RMSE values). These findings validate
and boosting rounds for XGBoost were optimized. The the choice of linear regression as the second-level learner,
FCN model used 200 iterations with Kaiming initializa- which enhances interpretability and optimally aggregates
tion and the Adam optimizer, while other models followed the outputs of base learners.
their respective GA-optimized configurations to achieve the
best performance on the validation set. To minimize the D. ENGINEERING TEST -NORMAL OPERATION
impact of random fluctuations, each model was trained The previous section demonstrated the effectiveness and
10 times independently. Table 5 summarizes the perfor- superiority of the proposed ensemble learning framework.
mance of these models on the test dataset, with the best On one hand, the heterogeneous integration strategy ensures
results highlighted. Across 10 independent training runs, the sufficient diversity among the base models, meeting the
Stacking model consistently achieved better average root requirements for ensemble learning. On the other hand, the
mean square error (RMSE) compared to other mainstream proposed model exhibits superior performance compared to
frameworks. Notably, the Stacking model demonstrated the other state-of-the-art regression models. In this section, the
smallest maximum RMSE among all models, reflecting its feasibility of the proposed Stacking model will be validated
superior performance in controlling the upper limit of pre- in a practical engineering context.
diction errors. The analysis focuses on two in-service power transformers
at a steel plant with electric arc furnaces: a 50 MVA/110 kV
TABLE 6. Performance comparison. transformer and a 114 MVA/220 kV transformer. Due to the
specific working conditions of the steel plant, the neutral
points of these transformers are grounded without neutral
point isolation devices, resulting in the presence of DC com-
ponents on the secondary side. Additionally, the transformers
experience significant load fluctuations, providing an oppor-
tunity to measure the effects of varying operational capacities
on transformer vibration performance.

The results in Table 5 highlight the advantages of the

proposed Stacking model (Stacking) over other frameworks,
with key observations summarized as follows:
a) The Stacking model achieved the lowest average RMSE
(2.018) across 10 independent training runs, outperform-
ing alternative frameworks, including Stacking-2 (2.173)
and Stacking-3 (2.094), as well as XGBoost (2.104). This
demonstrates the effectiveness of the proposed model
in consistently delivering accurate predictions, benefit- FIGURE 11. Prediction errors of transformer A.
ing from the complementary strengths of diverse base
learners. Figure 11 and 12 depicted the mean prediction errors
b) The Stacking model maintained a low maximum RMSE at measurement points under different operating conditions
(2.312), indicating its robustness in minimizing prediction for the two transformers (transformer A with 50 MVA/
errors even under challenging conditions. In comparison, 110 kV electrical specification and transformer B with
models like Stacking-3 (2.507) and FCN (2.501) exhibited 114 MVA/220 kV electrical specification). The results
higher maximum RMSE values, suggesting greater vari- include comparisons of predicted vibration values with actual
ability and reduced reliability. measurements across varying load conditions. These Figures
c) Traditional machine learning models, such as DTR (aver- provide a comprehensive evaluation of the model’s perfor-
age RMSE: 2.677) and SVM (average RMSE: 2.513), mance in capturing the complex vibrational behaviors of
showed significantly higher prediction errors, reinforcing power transformers under real-world operating conditions.
the advantage of the heterogeneous ensemble strategy The Stacking model demonstrates robust overall predic-
employed by the Stacking framework. tion performance, as evidenced by the generated surface

37824 VOLUME 13, 2025

B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

VI. CONCLUSION
This study presents a comprehensive framework for vibration
prediction in power transformers, with the key conclusions
summarized as follows:
(a) A parametric finite element model combined with the
Sobol sequence was used to construct a representative
dataset for transformer vibration analysis. Global sen-
sitivity analysis was conducted to quantify the variance
contributions of design variables, providing insights into
the key factors influencing transformer vibration.
(b) A Stacking ensemble learning framework was developed,
integrating DTR, ELM, and SVM as first-level models,
with penalized linear regression (LR) as the second-level
learner. This heterogeneous ensemble strategy effectively
FIGURE 12. Prediction errors of transformer B.
improved prediction accuracy and robustness by leverag-
ing the strengths of diverse base models.
plot of absolute prediction errors in Figure 11; however, the (c) The proposed model was validated using five-fold cross-
prediction errors show a tendency to increase with higher validation, heuristic hyperparameter optimization, and
DC components and operational capacities, primarily due engineering tests on transformers with varying electrical
to two reasons: first, the intensified nonlinear magnetization specifications. The results demonstrated the model’s gen-
process under DC bias significantly complicates the vibra- eralizability, adaptability, and reliability, showcasing its
tion characteristics of power transformers, thereby increasing potential for practical industrial applications.
the prediction difficulty; second, the theoretical model used
in this study cannot fully equivalently represent the inter- VII. LIMITATION AND FUTURE WORK
nal connection structures and secondary components of the This study has certain limitations and future work needs to be
transformer, which introduces modeling inaccuracies under conducted:
varying operational conditions; despite these challenges, (a) The combination of heuristic algorithms with five-fold
the proposed Stacking framework successfully mitigates cross-validation for model training results in relatively
major deviations by integrating heterogeneous base learners, high time and space complexity. Future work can focus
ensuring improved prediction accuracy across diverse opera- on simplifying the hyperparameter optimization pro-
tional scenarios and validating its effectiveness in addressing cess while maintaining controllable prediction errors to
complex prediction tasks. Differing from transformer A, improve computational efficiency.
transformer B features a unique design with a strong oil (b) The model could be further enhanced by incorporating
guide structure tailored to meet specific technical agreements, additional factors such as operating conditions, sec-
making it a non-standard product; therefore, analyzing the ondary component connections, and transformer service
errors for this transformer provides valuable insights into the years to enrich the feature set and improve prediction
generalization ability of the ensemble learning model. In con- accuracy. These improvements will further extend the
junction with Figure 12 It can be observed that the vibration applicability and robustness of the proposed framework.
prediction model retains a certain degree of generality, (c) Further research is required to investigate the influ-
achieving overall controllable prediction errors with a max- ence of bolt connections on transformer vibration across
imum error of less than 3 dB; moreover, the error response different capacity levels. Bolt connections introduce
surface for transformer B closely aligns with that of trans- additional structural complexity, which may have a sig-
former A, showing a similar trend where absolute prediction nificant impact on vibration characteristics, especially for
errors increase with higher DC components and operational large-capacity power transformers. Future studies should
capacities. focus on developing precise models that include bolt
The proposed Stacking ensemble learning model demon- preloading
strates significant advantages in engineering applications,
offering excellent predictive accuracy for standard trans- REFERENCES
former models while maintaining robust generalizability for [1] M. V. Czernorucki, M. B. C. De Salles, S. L. Avila, F. A. Sobrinho,
novel structures. Its ability to adapt to unknown designs, W. W. Calil, and J. R. Cardoso, ‘‘Multi-objective design optimiza-
tion for HVDC-LCC converter transformers: Analytical and FEA-
such as non-standard products with unique structural features, based comparison,’’ IEEE Access, vol. 11, pp. 23032–23045, 2023, doi:
ensures reliable performance with controllable prediction 10.1109/ACCESS.2023.3253392.
errors, highlighting its potential as a versatile and effec- [2] E. I. Amoiralis, M. A. Tsili, D. G. Paparigas, and A. G. Kladas, ‘‘Global
transformer design optimization using deterministic and nondeterministic
tive tool for vibration prediction across diverse transformer algorithms,’’ IEEE Trans. Ind. Appl., vol. 50, no. 1, pp. 383–394, Jan. 2014,
configurations. doi: 10.1109/TIA.2013.2288417.

VOLUME 13, 2025 37825

 B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

[3] IEEE Approved Draft Standard for General Requirements for Liquid- [21] X. Miao, P. Jiang, F. Pang, Y. Tang, H. Li, G. Qu, and J. Li, ‘‘Numerical
Immersed Distribution, Power, and Regulating Transformers, Standard analysis and experimental research of vibration and noise characteristics
PC57.12.00/D2.2, Aug. 2021, pp. 1–70. of oil-immersed power transformers,’’ Appl. Acoust., vol. 203, Feb. 2023,
[4] P. Jiang, Z. Zhang, J. Zhang, B. Deng, J. Deng, and Z. Pan, ‘‘Research Art. no. 109189, doi: 10.1016/j.apacoust.2022.109189.
on vibration characteristics and multi-parameter state recognition of [22] R. G. L. Koh, M. Ribeiro, L. Jabban, B. Fang, K. Nesovic, S. Bayat,
±500 kV converter transformer under fluctuating conditions,’’ Int. J. and B. W. Metcalfe, ‘‘A scoping review of machine learning applied
Electr. Power Energy Syst., vol. 136, Mar. 2022, Art. no. 107748, doi: to peripheral nerve interfaces,’’ IEEE Trans. Neural Syst. Rehabil. Eng.,
10.1016/j.ijepes.2021.107748. vol. 32, pp. 3689–3698, 2024, doi: 10.1109/TNSRE.2024.3468995.
[5] E. Stano and M. Kaczmarek, ‘‘Analytical method to determine the [23] K. S. Bohnsack, M. Kaden, J. Abel, and T. Villmann, ‘‘Alignment-free
values of current error and phase displacement of inductive current sequence comparison: A systematic survey from a machine learning
transformers during transformation of distorted currents higher har- perspective,’’ IEEE/ACM Trans. Comput. Biol. Bioinf., vol. 20, no. 1,
monics,’’ Measurement, vol. 200, Aug. 2022, Art. no. 111664, doi: pp. 119–135, Jan. 2023, doi: 10.1109/TCBB.2022.3140873.
10.1016/j.measurement.2022.111664. [24] A. Taufiqurrahman, A. G. Putrada, and F. Dawani, ‘‘Decision tree regres-
[6] M. Qian, F. Yin, Y. Yuan, P. Jiang, L. Wang, and L. Zhao, ‘‘Vibration sion with AdaBoost ensemble learning for water temperature forecasting
characteristics of ±800 kV converter transformers, Part—II: Under load in aquaponic ecosystem,’’ in Proc. 6th Int. Conf. Interact. Digit. Media
conditions,’’ Int. J. Electr. Power Energy Syst., vol. 159, Aug. 2024, (ICIDM), Dec. 2020, pp. 1–5, doi: 10.1109/ICIDM51048.2020.9339669.
Art. no. 110026, doi: 10.1016/j.ijepes.2024.110026. [25] S. Chatterjee, R. K. Gatla, P. Sinha, C. Jena, S. Kundu, B. Panda, L. Nanda,
[7] C. Hao, D. Xinyu, L. Ling, and W. Cai, ‘‘Test and analysis of acoustic and and A. Pradhan, ‘‘Fault detection of a Li-ion battery using SVM based
vibration characteristics of transformer in winding and iron core loose,’’ machine learning and unscented Kalman filter,’’ Mater. Today, Proc.,
in Proc. IEEE 3rd Int. Conf. Power, Electron. Comput. Appl. (ICPECA), vol. 74, pp. 703–707, Jan. 2023, doi: 10.1016/j.matpr.2022.10.279.
Jan. 2023, pp. 1644–1647, doi: 10.1109/ICPECA56706.2023.10075954. [26] K. H. Reddy, ‘‘Adaptive extreme learning machine using soft com-
[8] S. Yadav and R. K. Mehta, ‘‘Modelling of magnetostrictive vibration and puting fuzzy propositions—Validating operating state of solar energy
acoustics in converter transformer,’’ IET Electric Power Appl., vol. 15, system,’’ Appl. Soft Comput., vol. 164, Oct. 2024, Art. no. 111966, doi:
no. 3, pp. 332–347, Mar. 2021, doi: 10.1049/elp2.12025. 10.1016/j.asoc.2024.111966.
[9] Z. Tang and Z. Wang, ‘‘Study on statistical characteristics of audible noise [27] J. E. Rowe, M. D. Vose, and A. H. Wright, ‘‘Reinterpreting no free
and vibration signals of UHV transformers and reactors,’’ in Proc. IEEE lunch,’’ Evol. Comput., vol. 17, no. 1, pp. 117–129, Mar. 2009, doi:
Int. Conf. High Voltage Eng. Appl. (ICHVE), Sep. 2020, pp. 1–4, doi: 10.1162/evco.2009.17.1.117.
10.1109/ICHVE49031.2020.9279676. [28] N. S. Kovach and K. Littlejohn, ‘‘No free lunch with open mission
[10] H. Zhou, K. Hong, H. Huang, and J. Zhou, ‘‘Transformer winding systems,’’ in Proc. IEEE Nat. Aerosp. Electron. Conf., Aug. 2021,
fault detection by vibration analysis methods,’’ Appl. Acoust., vol. 114, pp. 233–238, doi: 10.1109/NAECON49338.2021.9696311.
pp. 136–146, Dec. 2016, doi: 10.1016/j.apacoust.2016.07.024. [29] W. Cui, C. Wan, and Y. Song, ‘‘Ensemble deep learning-based non-
crossing quantile regression for nonparametric probabilistic forecasting
[11] J. Zheng, J. Pan, and H. Huang, ‘‘An experimental study of wind-
of wind power generation,’’ IEEE Trans. Power Syst., vol. 38, no. 4,
ing vibration of a single-phase power transformer using a laser
pp. 3163–3178, Jul. 2023, doi: 10.1109/TPWRS.2022.3202236.
Doppler vibrometer,’’ Appl. Acoust., vol. 87, pp. 30–37, Jan. 2015, doi:
10.1016/j.apacoust.2014.06.012. [30] J. Lee, W. Wang, F. Harrou, and Y. Sun, ‘‘Wind power predic-
tion using ensemble learning-based models,’’ IEEE Access, vol. 8,
[12] Z. Li, X. Lang, B. Yang, X. Liu, H. Wang, and Z. Li, ‘‘Vibration and noise
pp. 61517–61527, 2020, doi: 10.1109/ACCESS.2020.2983234.
mechanism of a 110 kV transformer under DC bias based on finite ele-
[31] Y. He, H. Zhang, Y. Dong, C. Wang, and P. Ma, ‘‘Residential net load
ment method,’’ Global Energy Interconnection, vol. 7, no. 4, pp. 503–512,
interval prediction based on stacking ensemble learning,’’ Energy, vol. 296,
Aug. 2024, doi: 10.1016/j.gloei.2024.08.012.
Jun. 2024, Art. no. 131134, doi: 10.1016/j.energy.2024.131134.
[13] P. Jiang, Z. Zhang, Z. Dong, Y. Wu, R. Xiao, J. Deng, and Z. Pan,
[32] F. Kabakcioglu and E. Bayraktarkatal, ‘‘SOBOL sensitivity
‘‘Research on distribution characteristics of vibration signals of ±500
analysis and acoustic solid coupling approach to underwater
kV HVDC converter transformer winding based on load test,’’ Int. J.
explosion,’’ Ocean Eng., vol. 281, Aug. 2023, Art. no. 114752, doi:
Electr. Power Energy Syst., vol. 132, Nov. 2021, Art. no. 107200, doi:
10.1016/j.oceaneng.2023.114752.
10.1016/j.ijepes.2021.107200.
[33] C. R. Harris et al., ‘‘Array programming with NumPy,’’ Nature, vol. 585,
[14] C. Pan, C. Wang, Z. Liu, and X. Chen, ‘‘Winding vibration analysis of no. 7825, pp. 357–362, Sep. 2020, doi: 10.1038/s41586-020-2649-2.
unbalanced transformer based on electromagnetic-mechanical coupling,’’
Int. J. Electr. Power Energy Syst., vol. 134, Jan. 2022, Art. no. 107459, doi:
10.1016/j.ijepes.2021.107459.
[15] L. Zhu, H.-S. Yoon, H.-J. Cho, D.-J. Um, and C.-S. Koh, ‘‘Finite-
element analysis of magnetostriction force in power transformer based BAIDI SHI was born in Changzhou, Jiangsu,
on the measurement of anisotropic magnetostriction of highly grain-
China, in 1996. He received the M.S. degree from
oriented electrical steel sheet,’’ IEEE Trans. Magn., vol. 52, no. 3, pp. 1–4,
Hohai University, Changzhou, in 2021. He is cur-
Mar. 2016, doi: 10.1109/TMAG.2015.2481466.
rently pursuing the Ph.D. degree with the College
[16] W. Gong, Z. Zhang, R. Hou, H. Wang, Z. Xu, A. Lin, J. He, W. Fan, and
J. Wang, ‘‘Magnetostriction and the influence of harmonics in flux density of Mechanical and Electrical Engineering, Hohai
in electrical steel,’’ IEEE Trans. Magn., vol. 51, no. 11, pp. 1–4, Nov. 2015, University. His current research interests include
doi: 10.1109/TMAG.2015.2440320. multi-physics simulation, machine learning, and
[17] H. Qiang, N. Jingkai, Z. Songyang, X. Weimin, J. Shengchang, and multi-objective optimization.
C. Xin, ‘‘Study of transformer core vibration and noise generation
mechanism induced by magnetostriction of grain-oriented silicon steel
sheet,’’ Shock Vibrat., vol. 2021, no. 1, Jan. 2021, Art. no. 8850780, doi:
10.1155/2021/8850780.
[18] X. Liu, C. Sun, Y. Wang, F. Jiang, and C. Zhang, ‘‘Vibration character-
istic analysis of transformers influenced by DC bias based on vibration YONGFENG JIANG was born in Shanxi, Xi’an,
half-wave energy method,’’ Int. J. Electr. Power Energy Syst., vol. 128, China, in 1974. He received the Ph.D. degree in
Jun. 2021, Art. no. 106725, doi: 10.1016/j.ijepes.2020.106725. material processing engineering and science from
[19] L. Li, D. Zhang, Z. Wang, Y. Zhang, X. Fan, and Y. Zhou, ‘‘Novel field- Shanghai Jiao Tong University, in 2004. He is
circuit assisted FEA of 110 kV power transformer for noise control and currently working as a Professor with the College
vibration reduction,’’ Int. J. Appl. Electromagn. Mech., vol. 64, nos. 1–4, of Mechanical and Electrical Engineering, Hohai
pp. 289–298, Dec. 2020. University. His current research interests include
[20] B. Lee, K. Lee, C. Park, S. Ryu, and J. Chung, ‘‘Load noise prediction of a multi-physics simulation, machine learning, and
power transformer,’’ J. Vibrat. Control, vol. 28, nos. 23–24, pp. 3719–3727, multi-objective optimization.
Dec. 2022, doi: 10.1177/10775463211036816.

37826 VOLUME 13, 2025

B. Shi et al.: Power Transformer Vibration Analysis Model Based on Ensemble Learning Algorithm

WEI XIAO was born in Jiangxi, Ganzhou, China, ZIXING LI was born in Zhongxiang, Hubei, China, in 1974. He received the
in 1994. He received the bachelor’s degree in trans- bachelor’s degree in transformer and reactor design and development. He
former design and simulation calculation from is currently a Senior Engineer with Changzhou XD Transformer Company
Hebei University of Technology. He is currently Ltd. His current research interests include transformer and reactor design and
working as an Engineer with Changzhou Xidian development.
Transformer Company Ltd. His current research
interests include transformer design, simulation,
and optimization.

JINGYU SHANG was born in Anyang, Henan,

China, in 1987. She received the bachelor’s degree
in mechanical design, manufacturing, and automa-
tion and the master’s degree in mechanical engi-
neering from Henan Polytechnic University, in
2011 and 2014, respectively. She is currently
pursuing the Ph.D. degree in water machinery
program with the College of Mechanical and Elec-
trical Engineering, Hohai University.

XINFU CHEN was born in Changzhou, Jiangsu, China, in 1977. He received

the bachelor’s degree in transformer design. He is currently working as a
MENG LI was born in Yancheng, Jiangsu, China, in 1974. He is currently a Senior Engineer with Changzhou Xidian Transformer Company Ltd. His
Senior Engineer with Changzhou XD Transformer Company Ltd. His current current research interest includes transformer design and development.
research interests include transformer design, siulation, and manufacturing.

VOLUME 13, 2025 37827