IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO.

4, APRIL 2021 4633

An Overview of Artificial Intelligence Applications

for Power Electronics
Shuai Zhao , Member, IEEE, Frede Blaabjerg , Fellow, IEEE, and Huai Wang , Senior Member, IEEE

Abstract—This article gives an overview of the artificial intel-

ligence (AI) applications for power electronic systems. The three
distinctive life-cycle phases, design, control, and maintenance are
correlated with one or more tasks to be addressed by AI, including
optimization, classification, regression, and data structure explo-
ration. The applications of four categories of AI are discussed,
which are expert system, fuzzy logic, metaheuristic method, and
machine learning. More than 500 publications have been reviewed
to identify the common understandings, practical implementation
challenges, and research opportunities in the application of AI
for power electronics. This article is accompanied by an Excel file
listing the relevant publications for statistical analytics.
Index Terms—Artificial intelligence (AI), design, intelligent
controller, power electronic systems, predictive maintenance,
prognostics and health management (PHM).
Fig. 1. Annual number of publications of AI in power electronics since 1990.
I. INTRODUCTION
OWADAYS artificial intelligence (AI) is expanding
N rapidly and is one of the most salient research areas
during the past several decades [1], [2]. The aim of AI is to
Meanwhile, the rapid development of data science, including
sensor technology, Internet-of-Things (IoT), edge computing,
digital twin [9], and big data analytics [10], [11], provides a
facilitate systems with intelligence that is capable of humanlike wide variety of data for power electronic systems throughout
learning and reasoning. It possesses tremendous advantages different phases of its life-cycle. The increasing volume of
and has been successfully applied in numerous industrial areas, data enables immense opportunities and lays a solid foundation
including image classification, speech recognition, autonomous for the AI in power electronics. AI is able to exploit data to
cars, computer vision, etc. With immense potentials, power elec- improve product competitiveness by global design optimization,
tronics benefit from the development of AI. There are various intelligent control, system health status estimation, etc. As a
applications, including design optimization of power module result, the research in power electronics can be conducted from
heatsink [3], intelligent controller for multicolor light-emitting a data-driven perspective, which is beneficial especially to com-
diode (LED) [4], maximum power point tracking (MPPT) con- plex and challenging cases.
trol for wind energy conversion systems [5], [6], anomaly detec- Due to the specific challenges and characteristics of power
tion for inverter [7], remaining useful life (RUL) prediction for electronic systems, e.g., high tuning speed in control, high
supercapacitors [8], etc. By implementing AI, power electronic sensitivity in condition monitoring for aging detection, etc.,
systems are embedded with capabilities of self-awareness and the implementation of AI in power electronics has its own
self-adaptability, and therefore, the system autonomy can be features that are different from other engineering areas, e.g.,
improved. image classification. Therefore, there is a pressing need for an
overview of AI in power electronics to expedite synergy research
Manuscript received June 4, 2020; revised August 6, 2020; accepted Septem-
ber 11, 2020. Date of publication September 18, 2020; date of current version and interdisciplinary applications. Based on literature review, in
November 20, 2020. This work was supported in part by the Innovation Fund this article, the applications of AI in power electronics are cate-
Denmark through the project of Advanced Power Electronic Technology and gorized into three aspects, i.e., design, control, and maintenance.
Tools, and in part by the Villum Foundation through the project of Light-AI
for Cognitive Power Electronics. Recommended for publication by Associate Fig. 1 shows the annual number of publications related to AI
Editor Prof. Kyo-Beum Lee. (Corresponding author: Huai Wang.) for power electronics since 1990. The statistical data are based
The authors are with the Department of Energy Technology, Aalborg on searching the IEEE Xplore from the journals IEEE TRANSAC-
University, 9220 Aalborg, Denmark (e-mail: szh@et.aau.dk; fbl@et.aau.dk;
hwa@et.aau.dk). TIONS ON POWER ELECTRONICS, IEEE JOURNAL OF EMERGING
This article has supplementary downloadable material available at https:// AND SELECTED TOPICS IN POWER ELECTRONICS, IEEE TRANS-
ieeexplore.ieee.org, provided by the authors. ACTIONS ON INDUSTRIAL ELECTRONICS, IEEE TRANSACTIONS
Color versions of one or more of the figures in this article are available online
at https://ieeexplore.ieee.org. ON INDUSTRIAL INFORMATICS, and IEEE TRANSACTIONS ONIN-
Digital Object Identifier 10.1109/TPEL.2020.3024914 DUSTRY APPLICATIONS. The data of 2020 are up to May 2020.

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
4634 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

As a result, a total of 444 relevant journal papers are identified, 2) A timeline map is provided to illustrate the milestones of
which can be found in the supplemental Excel file. It can be seen AI algorithms and power electronic applications. More-
that the implementations of AI have been drastically increased over, it presents the quantitative information of the method
and experienced a spectacular dynamism over the last few years. usage percentages and application trend.
The number of publications for control is continuously increas- 3) The advantages and limitations of AI algorithms are
ing and it is the most active research area. Since 2007, there is comprehensively investigated. Exemplary applications are
an increase regarding the design and maintenance applications, provided for AI in each life-cycle stage, where the chal-
and such trends are more evident in the last two years. lenges and future research directions are discussed.
It is found that several existing reviews in the literature The rest of this article is organized as follows. Section II
are related to this topic. In [12], the metaheuristic methods presents the functions, methods, and milestones of AI in power
for stochastic optimization for power quality and waveform, electronics. The applications of AI in design, control, and main-
circuit design, and control tuning are reviewed. It focuses on tenance are discussed in Sections III–V, respectively. The out-
the optimization tasks only. The details of neural network (NN) look on the AI applications for power electronics is put forward
in industrial applications are presented in [13] with the design in Section VI. Finally, Section VII concludes this article.
of network structure, training methods, and application consid-
erations. It covers a broad scope of engineering applications
beyond power electronics. In [14], a comprehensive review is II. FUNCTIONS AND METHODS OF AI FOR POWER
given on the applications of NN in power electronics. Several ELECTRONIC SYSTEMS
specific examples of control and system identification are de- Fig. 2 gives a summary of the methods, functions, and ap-
tailed. Nevertheless, other AI techniques, such as fuzzy logic, plications of AI for power electronics. It can be seen that AI
metaheuristic methods, etc., have not been discussed. Although has been extensively applied to the three distinctive life-cycle
these techniques are further discussed in [15], it emphasizes on phases of power electronic systems, including design, control,
illustrative examples while an in-depth analysis of AI algorithms and maintenance.
is not provided. In [16], an intensive discussion of metaheuristic As a functional layer between AI and power electronic appli-
methods for MPPT in photovoltaic (PV) systems is presented. cations, the essential functions of AI are categorized as optimiza-
In [17], the AI techniques applied to PV systems are reviewed, tion, classification, regression, and data structure exploration.
which focus on the specific PV applications only. 1) Optimization: It refers to find an optimal solution max-
Maintenance [18] in power electronics is a topic that includes imizing or minimizing objective functions from a set
reliability, condition monitoring, RUL prediction, etc. Several of available alternatives given constraints, equalities, or
review papers in the past decade can be found in [19]–[22]. inequalities that the solutions have to satisfy. For example,
In [19], a state-of-the-art analysis of the condition monitoring in the design task, optimization serves as a tool to explore
and fault detection in power electronics is presented. However, an optimal set of parameters that maximize or minimize
it only includes a very limited AI-based fault detection method. design goals with design constraints.
In [20], a review of condition monitoring techniques for ca- 2) Classification: It deals with assigning input information or
pacitors in power electronic converters is presented. It includes data with a label indicating one of the k discrete classes.
only the AI-based parameter identification methods. In [21], Specifically, anomaly detection and fault diagnosis in
the methods in prognostics and health management (PHM) maintenance is a typical classification task to determine
of information and electronics-rich systems are summarized. fault labels with condition monitoring information.
This article only discusses the category of AI algorithms in the 3) Regression: By identifying the relationship between input
PHM area while there is no algorithm detail or comparative variables and target variables, the goal of regression is
analysis. In [22], machine learning methods applied in reliability to predict the value of one or more continuous target
management of energy systems are summarized. It focuses on variables given input variables. For example, an intelli-
the machine learning method and the maintenance task only. gent controller can be facilitated with a regression model
A tutorial [23] regarding “Artificial Intelligence Applications to between the input electrical signals and the output control
Power Electronics” is presented on the 2019 IEEE Energy Con- variables.
version Congress and Exposition. It serves as an introductory 4) Data Structure Exploration: It consists of data clustering
level presentation. Nevertheless, the desirable details of the AI that discovers groups of similar data within a dataset,
algorithms and their comparisons are not available. density estimation that determines the distribution of data
As a result, it lacks a comprehensive review of the AI algo- within the input space, and data compression that projects
rithms and applications for power electronics. From a life-cycle high-dimensional data down to low-dimensional data for
perspective, this article aims to fill this gap and comprehensively feature reduction. For example, in maintenance, the degra-
review the published research in power electronics using AI dation state clustering is within the data structure explo-
techniques, which needs a systematic consolidation. The con- ration category.
tributions of this article include the following. According to the surveyed 444 relevant journal papers, Fig. 3
1) The AI algorithms in power electronics are systematically shows a Sankey diagram of application usage statistics of AI
investigated from a life-cycle perspective, where the re- methods in the life-cycle of power electronic systems. Specif-
lationships of the relevant AI algorithms, their essential ically, the percentages of AI application in the design, control,
functions, and the relevant applications are identified. and maintenance are 9.8%, 77.8%, and 12.4%, respectively.
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4635

Fig. 2. Application of AI in the life-cycle of power electronic systems. Section II-A implies that the relevant discussions are presented in part A of Section II.

expert knowledge in a Boolean logic catalog, based on which the

IF–THEN rules in human brain reasoning are simulated. It is an
intelligent system simulating the inference process that answers
the why-and-how inquires based on the database. The database
is from either field expert experience or simulation data, facts,
and statements. It can be continuously updated. The technical
details of expert system are given in [17], and several exemplary
applications can be found in [15] and [28].
It is worth mentioning that the applications of expert system
are as low as 0.9% according to the usage statistics in Fig. 3. It
is because the expert system is generally based on system prin-
ciples and rules, which relates strongly to the system of interest
and lacks universality. It applies to well-defined domains only
Fig. 3. Sankey diagram of AI methods and applications in each phase of the
with solid expert rules. Besides, due to the rapid development
life-cycle of power electronic systems. The statistical usages and percentages of computational platforms, the functions of expert system can
are based on the data in Fig. 1. be replaced with other advanced AI methods (e.g., fuzzy logic
and machine learning) with superior capabilities in inference
and approximation.
Regarding the functions, the percentages of optimization, clas-
sification, regression, and data structure exploration are 33.3%,
6.6%, 58.4%, and 1.7%, respectively. It shows that most of the B. Fuzzy Logic
tasks of AI in power electronics are essentially regression and
Similar to expert system, fuzzy logic is also a rule-based
optimization. The AI methods can be generally categorized as
method while it extends the Boolean logic into a multivalued
expert system, fuzzy logic, metaheuristic methods, and machine
case. Fuzzy logic is an ideal tool to tackle system uncertainties
learning. Their application percentages are 0.9%, 21.3%, 32.0%,
and noisy measurements [29]–[31]. Instead of using the precise
and 45.8%, respectively. It suggests that the largest portion of
input crisp value directly, fuzzification is first performed with
AI in power electronics is with the machine learning. These
the fuzzy sets consisting of several membership functions to a
methods will be detailed subsequently. Note that a comprehen-
range of 0–1. The fuzzy input signals are then aggregated with
sive but still not exhaustive investigation is conducted. Only the
fuzzy rules in the inference step. Defuzzification is subsequently
relevant AI methods that are widely applied to power electronics
performed on the inference result by considering the degree
are considered.
of fulfillment and output a crisp value. As a result, the crisp
value is manipulated in a fuzzy space that completes nonlinear
A. Expert System
mapping between the input and output with elaborately designed
Expert system is the earliest method in AI that is effec- principles.
tively implemented in industrial applications [17]. The expert In most applications, a fuzzy logic method mainly consists
system [24]–[27] is essentially a database that integrates the of four parts [30]: fuzzification, rule inference, knowledge base,
4636 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

and defuzzification. First, fuzzification is performed on the input

of linguistic variables with membership functions, including tri-
angular, trapezoidal, Gaussian, bell-shaped, singleton, and other
customized shapes. Second, the inference module integrates the
signals together according to IF–THEN fuzzy rules in the knowl-
edge base derived from expert experience. Third, defuzzification
is performed on the signal for output. One example of the fuzzy
rule is
Antecedent: IF X is Medium AND Y is Zero,
Consequent: Then Z is Positive.
For both the antecedent and consequent, the degree of fulfill-
ment is determined by the membership functions. The type of Fig. 4. Usage statistics of population-based metaheuristic methods in opti-
fuzzy inference scheme is categorized as Mamdani-type [30], mization of power electronics. The statistical results are obtained based on the
data in Fig. 1.
[32]–[35] and Takagi–Sugeno–Kang-type (TSK-type) [31],
[36]–[38]. For the Mamdani-type fuzzy inference scheme, the
membership functions of the antecedent and the consequent are exploration stage includes only one candidate solution and it
shape-based functions, e.g., triangular. For the TSK-type fuzzy evolves into another solution according to a certain rule. The
inference scheme, the membership function of the antecedent performance of this method is mainly based on the quality and
part is identical to the Mamdani-type while that of the conse- efficiency of the rule. As a result, the convergence speed of
quent is singleton at several constant values. Typically, more the trajectory-based methods is generally slow and the final
fuzzy sets are needed for the Mamdani-type scheme compared solution is prone to local rather than global solution for non-
to the TSK-type scheme for the same task. Compared to the convex optimization tasks. For the population-based methods,
fuzzy terms in the Mamdani-type, the membership function in multiple candidate solutions are randomly generated. At each
the TSK-type scheme can be functional type as either linear iterative exploration, these candidate solutions are diversified
or constant, which is more powerful and accurate in nonlinear (e.g., crossover in the GA) or incorporated and replaced with
approximation. More theoretical details of fuzzy logic are dis- new candidate solutions to improve the quality of the popu-
cussed in [15], [39]. lation at the present generation. As a result, the suitability of
Note that expert experience plays a critical role in the design of the population is iteratively improved to approach the optimal
the membership function and the fuzzy rule, and such a method solution. Compared to the trajectory-based methods, they are
is applicable to experts only in most cases. From this perspective, superior in the convergence speed, the global searching capa-
the prior information and expert experience can be coped with bility, and especially useful for large-scale optimization tasks.
fuzzy logic and then incorporated with other AI techniques as a Nevertheless, the computational burden of the population-based
hybrid method. methods is more intensive. This challenge needs to be considered
for online application cases where efficiency and speed are of
most significance. Table I shows a summary of the metaheuristic
C. Metaheuristic Methods
methods in the area of power electronics with their advantages
Once the optimization task of specific applications is formu- and limitations. These metaheuristic methods are qualitatively
lated, the optimal solution can be obtained by either a determinis- compared in terms of several critical features, including imple-
tic programming method (e.g., linear or quadratic programming) mentation simplicity, global convergence, convergence speed,
or a nondeterministic programming method, i.e., metaheuristic and parallel capability.
method. The deterministic programming methods need to calcu- Due to enormous advantages, most of the optimization tasks in
late the gradient and Hessian matrices [40], which is challenging power electronics are solved with the population-based methods.
for most of the optimization tasks in power electronics due to the It can be seen from Table I that there are various population-
complexity. Metaheuristic methods serve as a general end-to-end based methods with the improved variants for optimization tasks
tool that needs less expert experience and is efficient and scalable in power electronics. They are developed and improved with dif-
for various optimization tasks. ferent biological inspirations. In addition to the earlier widely ap-
The metaheuristic methods [12] are generally developed plied metaheuristic methods, several other emerging approaches
with inspirations of biological evolution, e.g., genetic algorithm have been applied in a limited scale, e.g., biogeography-based
(GA) [41] by process of natural selection, ant colony optimiza- optimization [72], crow search algorithm [73], grey wolf op-
tion (ACO) algorithm [42] by simulating ants in finding an timization [74], firefly optimization algorithm [16], bee al-
efficient path for foods. The exploration of optimal solution gorithm [75], colonial competitive algorithm [76], teaching-
is motivated by the trial-and-error process. The metaheuristic learning-based optimization [77], etc. It is worth mentioning
methods can be categorized as trajectory-based methods (tabu that the selection of the best method is not a simple task, which
search method [43], simulated annealing method [44], etc.) and is application-dependent [12]. GA and PSO are the two most
population-based methods [GA, particle swarm optimization popular metaheuristic methods applied to power electronics, as
(PSO) [45], ACO, differential evolution [46], immune algo- shown in Fig. 4. They are the fundamentals and representatives
rithm (IA) [47], etc.]. For the trajectory-based methods, each for evolutionary algorithms and swarm intelligence algorithms,
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4637

TABLE I
APPLICATIONS OF METAHEURISTIC METHODS IN POWER ELECTRONICS

Superior: +++, intermediate: ++, inferior: +.

respectively, based on which various variants are developed. cases, its generalization on new inputs is one of the most critical
Practitioners can choose the method considering its superiority performance factors of supervised learning methods.
according to Table I. Generally, supervised learning methods can be categorized
Note that there is no guarantee for a global optimum for meta- into connectionism-based methods (i.e., NN method), proba-
heuristic methods, but the solution is generally satisfactory and bilistic graphical methods, and memory-based methods (i.e.,
acceptable for most practical applications. For more theoretical kernel method). For NN methods, knowledge learned from the
details of the metaheuristic methods, readers can refer to [16] training dataset is facilitated and transferred as the connection
and [78]. weights and structures of the network. Numerous research has
been devoted to improving the performance of NN methods.
These improvements are from two aspects for applications in
D. Machine Learning
power electronics. The first aspect deals with enabling the
Machine learning is designed to automatically discover prin- uncertainty capability in handling the noisy signal of the NN
ciples and regularities with experience from either collected data to improve the method robustness. This feature is facilitated
or interactions by trial-and-error. For applications in power elec- by integrating the fuzzy logic into the NN as the fuzzy NN
tronics, it is categorized as supervised learning, unsupervised (FNN) or its variants (e.g., adaptive neurofuzzy inference system
learning, and reinforcement learning (RL). (ANFIS) [101]). The second aspect is for dynamic-performance
1) Supervised Learning: With the training dataset consisting improvement of the NN to tackle time-series dataset cases, e.g.,
of input-and-output pairs, the supervised learning aims to estab- intelligent controller, RUL prediction. Compared to the con-
lish the mapping and functional relationships between the inputs ventional NN where the network weights are independent, the
and outputs implicitly. This feature is especially useful for cases transient performance is facilitated by sharing weights between
in power electronics where system models are challenging to different layers and network cells. The weight sharing can be
formulate. Generally, the tasks of the supervised learning include implemented either in a shallow scale with a convolutional
classification and regression. For classification, its output of structure (e.g., 1-D convolutional NN (CNN), time-delayed NN
the input-and-output pairs in the training dataset deals with a (TDNN) [114]), or in full and deep scale by using a recurrent
finite number of discrete categories to be labeled. For example, unit as recurrent NN [105]. Generally, the modeling capability
the fault diagnosis for a multilevel inverter [94] is a typical of recurrent unit implementation is superior to the one with
classification task where the discrete fault label needs to be a convolutional structure. More theoretical details of the NN
identified given the input fault information. For a regression methods are discussed in [1, Ch. 5], [13], and [14].
task, the output of the input-and-output pairs consists of one or The probabilistic graphical methods obtain knowledge from
more continuous variables. An example of regression is the RUL the data by using a diagrammatic representation of input-and-
prediction of IGBTs [114] where the output, i.e., the residual output pairs. The diagrammatic representation implies the condi-
useful lifetime, is a continuous variable. Once the model is tional dependence relationship between the decision variables.
trained, it is ready to evaluate new data points that differ from The underlying relationship in the model is formulated in the
the training dataset. The model capability in dealing with new Bayesian framework [1] and can be inferred in a probabilistic
data points, i.e., the ones in the testing dataset, is termed as way. Thus, the interpretability of the model is much better
the generalization. Since the training dataset comprises only compared to NN methods. Besides, the probabilistic graphical
a limited amount of possible input-and-output pairs in most model is superior in dealing with uncertainty and incomplete
4638 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

TABLE II
SUPERVISED LEARNING METHODS AND THE APPLICATIONS TO POWER ELECTRONICS

knowledge. One of the typical probabilistic graphical methods burden, sparse solutions are proposed as SVM and relevance
is the Bayesian network [117]. More theoretical details of the vector machine (RVM), where the parameter estimation is im-
probabilistic graphical methods are given in [1, Ch. 8]. proved based on Bayesian methods. With the sparse solution,
For the NN methods and the graphical methods, the training only a subset of the training dataset is applied to the testing
dataset is discarded when the training is completed. While the stage, and thus, it is more efficient compared to the conventional
training dataset in kernel methods is kept and used in the testing kernel methods. More theoretical details of the kernel methods
stage, and the learned knowledge is facilitated as the identifi- are discussed in [1, Chs. 6 and 7]. Generally, the requirement
cation of critical data points (e.g., support vectors in support of the training dataset for the kernel methods is lower than the
vector machine (SVM) [126]) or subset in the training dataset. NN methods. Therefore, the kernel methods are more suitable
One typical kernel method is Gaussian processes, which has for the cases with a small dataset. While due to the training
been applied to the RUL prediction of IGBTs in [119]. Note dataset is needed in the testing stage, the memory requirement
that the conventional kernel methods (e.g., Gaussian processes) of the kernel methods is higher than the NN methods. The
are computationally intensive due to the whole training dataset is involvement of the training dataset also limits the speed per-
applied to the testing stage. To avoid the excessive computational formance at the testing stage. It should be considered for online
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4639

TABLE III
UNSUPERVISED LEARNING METHODS AND THE APPLICATIONS TO POWER ELECTRONICS

applications where the execution time is critical, e.g., control

application.
As a result, Table II shows a summary of the supervised
learning methods and their variants in power electronics, in terms
of the advantages, limitations, and exemplary applications.
2) Unsupervised Learning: Compared to the supervised
learning where the dataset is input-and-output pairs, unsuper-
vised learning has no output data for the learning target during
the learning process. Generally, the tasks of unsupervised learn-
ing in applications of power electronics can be categorized as
data clustering and data compression.
For the data clustering, it explores the regularities from the
smeared dataset and partitions the dataset into several different
groups or clusters according to their similarities. In this way, Fig. 5. Usage statistics of machine learning methods in power electronic
the data characteristics within the same cluster are similar to systems. The statistical results are obtained based on the data in Fig. 1.
each other and different from the ones in other clusters. One
typical data clustering application is the identification of the
discrete health state from the continuous degradation data [131] future. The Q-table is an informative policy matrix that records
in the condition monitoring of power electronic converters. the optimal action to be taken given the particular condition
The purpose of the data compression is to eliminate excessive variables. More theoretical details of RL can be found in [139].
information in the dataset to reduce the number of features of One application example is the MPPT [5], [6], [140]. Note that
the dataset. For example, using principal component analysis RL obtains the experience from the interactions between systems
(PCA) [127], a reduced representation of the dataset is obtained instead of existing datasets. It is, thus, more favorable for the
with a much fewer number of features, which yet maintain the cases where the system is with less knowledge or its model is
integrity of the dataset. challenging to formulate.
Generally, these unsupervised learning algorithms serve as As a summary, Fig. 5 presents the usage statistics of the
the data-preprocessing before it goes to the subsequent analyt- machine learning methods. Supervised learning is dominantly
ics (e.g., fault diagnosis). Although this step is optional, it is applied to power electronics. The reason is that the supervised
beneficial to reducing the computational burden and improving learning is a versatile tool, which is typically the central part of
the analytics accuracy. Table III gives a summary of typical un- the majority of machine learning-related applications in power
supervised learning methods for power electronic applications. electronic systems.
More unsupervised learning methods and theoretical details can
be found in [137].
3) Reinforcement Learning: In contrast to the supervised E. Timeline of Relevant AI Methods and Applications in
learning and the unsupervised learning, RL does not require Power Electronics
a training dataset. Instead, it aims to find a suitable action Fig. 6 summarizes the milestones of the relevant AI methods
strategy that maximizing the reward for a specific task, which and their applications in power electronics. It includes the year
is essentially a dynamic programming or optimization task. when the algorithm is first proposed, the first application in
This goal-oriented strategy is formulated from interactions with power electronics, the milestones of relevant AI algorithms, and
systems or simulation models by a trial-and-error process [138]. applications in terms of each method. It should be noted that
In this way, it accumulates experience progressively and learns the information is to the best knowledge of the authors. Also,
a specific strategy that maximizes the predefined goal. Theoret- the timeline is not extensive to include all of the existing AI
ically, RL is a Markov decision process [139]. The training of algorithms. Instead, only the ones that show great potentials in
RL aims to develop a Q-table in terms of an action selection power electronics are included. According to Fig. 6, following
policy, which can maximize the total expected rewards over the can be noted.
4640 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

Fig. 6. Timeline of relevant AI methods and applications in power electronics. The milestones are identified considering the significant algorithm variants and the
relevant applications. It is organized as the form of (significant variants)-application-year. Significant variant is specifically indicated. Otherwise, it is a standard
algorithm.

1) The application of both expert system and fuzzy logic the potentials of NN methods in dealing with complex
is moderate nowadays, especially for the expert system. tasks in power electronic systems. Second, the structure
Before the 2000s, their practical implementations are de- of NN is quite flexible to incorporate other AI methods for
veloped in the presence of the limited performance of com- performance improvement, implying numerous method
puting hardware, which has been significantly improved variants.
to date. This rapid development of computing hardware 4) There is an increasing trend of applications with kernel
facilitates and accelerates the implementation of other methods and probabilistic graphical models. It is because
more powerful AI methods for replacing expert system most of these methods are formulated within the Bayesian
and fuzzy logic. framework, which possesses better generalization and in-
2) Metaheuristic methods are continuously evolving and ap- terpretability. Moreover, their computational burden can
plied to power electronics. They are used for a complete be well tackled with the platforms to date.
task or a key step jointly with other machine learning 5) RL is the latest frontier of the machine learning methods
methods. applied to power electronics, facilitated by the rapid de-
3) NN methods are the most active area for AI applications velopment of computing hardware.
for power electronics. The reason is twofold. First, the The following can be noted from Figs. 2, 3, and 6 about the
significant development of computing hardware unleashes comparisons for different AI methods.
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4641

1) Both metaheuristic methods and machine learning can 2) Constraint space: The constraint space defines feasible
be applied to optimization tasks. Specifically, machine- space, boundary, relationship, and limitation that the ob-
learning-based optimization (i.e., RL) focuses on the dy- jective function is subjected to. These constraints include
namic optimization involved with the decision-making either linear or nonlinear equalities and inequalities. They
(e.g., MPPT). Metaheuristic method is generally applied are derived from the practical design requirements, e.g.,
to the static optimization (e.g., heatsink design). geometry, volume, lifetime characteristics, cost, etc.
2) Both fuzzy logic and machine learning can be exploited for 3) Solution exploration: The defined optimization problem
classification tasks. Generally, machine learning is more is to maximize (or minimize) objective functions by ad-
accurate and flexible than fuzzy logic. justing the decision variables in the constraint spaces. AI
3) The regression task can be implemented with expert sys- methods, especially the metaheuristic methods, can be
tem, fuzzy logic, and machine learning. The implementa- applied to this step.
tion of expert system is simple but less powerful compared 4) Performance evaluation: The candidate solution can be
to fuzzy logic and machine learning. The implementation tested against the predefined objectives by using simula-
of fuzzy logic needs expert experience. Machine learning tion, hardwire-in-the-loop testing, prototype experiment,
is the most popular method and various algorithm variants etc. The results can be returned to previous steps for further
have been developed. It can be incorporated with fuzzy performance improvement and optimization.
logic for performance improvement. Instead of a sequential procedure, the design task is an iterative
4) Only machine learning can be applied to the task of data trial-and-error process. Based on the evaluation at each step,
structure exploration. the task may be reformulated, e.g., adjusting the objectives,
The following three sections discuss the applications of the modifying the constraint space, reconfiguring the programming
previously introduced AI methods in the design, control, and methods, etc. For conventional design in power electronics, it is
maintenance phases of power electronic systems, respectively. time-consuming and needs multiple iterative steps. For example,
the component alignment and the model selection rely on expert
experience and intuition without ample quantitative reference.
III. DESIGN In this way, the design performance will converge slowly to
Design in power electronics encompassing topology selec- the required standards. This drawback can be mitigated by AI
tion, component sizing, circuit synthesis, reliability considera- methods. They can be applied to 1) objective formulation for
tions, etc., is essentially an optimization task [145]. A typical the design time reduction, and 3) solution exploration for the
procedure for the design of power electronic systems comprises modeling and optimization.
following four steps.
1) Objective formulation: Objective functions are desirable
A. Design Time Reduction
design goals to be maximized or minimized. Generally,
the design goals in power electronics include component The formulation of design objective needs to be improved if
parameter [41], weight [146], volume [147], cost [146], its evaluation is computationally intensive. One application of
heatsink pattern [3], area [148], power loss [62], etc. It AI methods is a surrogate model in the objective formulation
is crucial for formulating the required or desired design to reduce the computational effort. The surrogate model yields
requirements to several explicit mathematical expressions an identical behavior to the system dynamics that are chal-
as a single objective, as given in (1), or multiple objectives, lenging to formulate or need intensive computational efforts to
as given in (2) [12], [145]: characterize. In the iterative design process, AI-based surrogate
model serves as a replacement that significantly reduces the
max f (x) (1) computational effort.
x
As an application of Design for Reliability (DfR), in [80],
max wT f (x), max f (x) two feed-forward NNs (FFNNs) are applied to the automated
x x
reliability design of power electronic systems. The first FFNN
s.t. g(x) ≤ 0, h(x) = 0, x ∈ [xl , xu ] (2) serves as a surrogate model emulating thermal characteristics
of power converters, by which the design parameters can be
where g(x) and h(x) are inequalities and equalities, re- mapped to the information of junction temperature variations.
spectively. xl and xu are the lower and the upper bound- Subsequently, the second FFNN is applied to map the annual
aries for decision variables x, respectively. Here, the max- mission profiles (e.g., annual solar irradiation and ambient
imization is the goal, which can simply be applied to the temperature) to the annual lifetime consumption. In this way,
minimization case. Note that, for multiple objectives in the nonlinear relationship between the designed parameters and
(2), it can be either solved by maximizing a scalar function the annual lifetime consumption is quantitatively characterized,
wT f (x) by weighting multiple objectives together or by which can accelerate the iterative design process.
optimizing objective vector f (x) directly, where Pareto Another example of AI for DfR of power electronic systems
front [62] can be applied to determine the optimal solution, is given in [109]. With superior capability in tackling time-series
e.g., the nondominated sorting GA method for multiobjec- data, a nonlinear autoregressive network with exogenous inputs
tive design optimization of power modules in [60]. (NARX) is applied to the thermal modeling of power electronic
4642 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

PV maximum power, PV maximum power point voltage, and

number of panels per string. The GA combining with the Pareto
front is applied to solve the multiobjective optimization task.
Besides, there is a specifically improved variant of GA for the
multiobjective optimization task, i.e., nondominated sorting GA
II (NSGA-II) [63].
In [45], the PSO is applied to the circuit synthesis of a power
Fig. 7. Nine different cell patterns for each blank cell [3]. A GA is applied electronic circuit, where the optimal values of components are
to determine the optimal combination of different cell patterns for the heatsink
design for minimizing the junction temperature.
explored to fulfill the design goals of better static and dynamic
performance. For this specific case, the simulation indicates that
the PSO yields a superior solution with less computational effort
systems considering the thermal cross-coupling effects. The compared to GA.
proposed NARX-based thermal model can be completed within In [70], the ACO is applied to determine the optimal compo-
around 109 s, which is a significant efficiency improvement com- nent values in a power electronic circuit, where the conventional
pared to the 1005 s of the conventional model. The error between ACO is extended to facilitate the optimization with continu-
the temperature estimated by the NARX-based thermal model ous component values and accelerate the optimization process.
and the actual measurement is less than 1 ◦ C. Experimental Moreover, the component tolerance is incorporated into the
results indicate that the NARX-based thermal model can replace optimization, which makes the proposed method more beneficial
the conventional model with less testing efforts and much less to practical applications.
computational burden. IV. CONTROL
In [79], considering the electrothermal interactions, an
FFNN is applied to construct the component behavior model Essentially, control applications with AI methods in power
of MOSFETs without any in-depth knowledge of the device electronic systems can be categorized as the optimization and
structure. Under the static state, the complicated nonlinear and the regression. Similar to the optimization in the design phase,
temperature-dependent characteristics between the variables, the optimization-related tasks in control applications are also
including drain-to-source voltage VDS , gate-to-source voltage dealing with metaheuristic methods. Several representative ap-
VGS , junction temperature Tj , and the output current ID are plications are given ahead.
established by using the NN. This compact model can drastically In [64], a GA is applied to the PID tuning of a programming
accelerate the design simulation process with a comparable logic controller, where the optimization goal is to minimize
accuracy. the error between the ideal step and ramp responses and the
ones initialized with proportional term Kp , integral term KI ,
and derivative term KD found by GA. Experimental analysis
B. Modeling and Optimization
indicates that the output performance of the optimized controller
The modeling and optimization of power electronic systems is is very close to the ideal step and ramp responses.
about specifying circuit topology, component model, component In [42], to overcome the challenges of multiple maximum
parameter, etc., such that system dimension, weight, operating power points in partially shaded situations for PV systems, an
frequency, etc., will result as optimal characteristics (e.g., power ACO-based MPPT method is proposed. It is compared with con-
loss, power density) given design constraints [12]. Specifically, ventional methods, including constant voltage tracking, perturb
the optimization method is applied to the solution exploration to & observe, and PSO. The experimental results indicate that the
provide an overall optimal configuration, where metaheuristic ACO-based MPPT method is superior in global convergence and
methods in AI can be exploited. As mentioned, the selection robustness to various shading patterns.
of a suitable metaheuristic method depends on the specific In [47], in a single-phase full-bridge inverter, an IA is applied
application. Several exemplary applications are given as follows. to find the optimal sinusoidal pulsewidth modulation (PWM)
In [3], GA is combined with finite-element analysis for the control sequences of four switches minimizing the total har-
automated heatsink design of a 50-kW three-phase inverter. As monic distortion (THD) of the output waveforms. The exper-
shown in Fig. 7, GA is applied to optimize the combination of iment indicates that the THD by using IA is 0.79%, which is
nine customized patterns to formulate a complex cell pattern superior to that of the conventional control method of hysteresis
of heatsink. The goal is to minimize the junction temperature current PWM with 1.23% and the GA solution with 0.99%.
of power semiconductor devices. Compared to the conventional Moreover, the IA is superior to the GA in convergence speed.
design with a regular cell pattern, the proposed method formu- More examples of optimization-related control applications can
lates a heatsink solution with 27% less in size and 6% lower in be found in [12].
junction temperature. The regression-related tasks in control applications are deal-
In [62], the design of a 500-kW solar power-based microgrid ing with the nonlinear mapping of system inputs and outputs
system is formulated as a multiobjective optimization task, in a static or dynamic way. Specifically, it is concerned with
which maximizes the average power distribution and mini- regulating systems to ensure intended performance output with
mizes the system weight simultaneously. It explores the optimal system principles. Several limitations of conventional methods
values of four microgrid parameters, including battery voltage, are identified, which are as follows.
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4643

Fig. 8. Fuzzy logic-based controller for a variable-speed wind generation system [30]. MFs: Membership functions. In the rule matrix table: P: positive; V: very;
B: big; M: medium; ZE: zero; N: negative.

1) The controller configuration requires in-depth knowledge is tuned by using the Lyapunov stability theorem to ensure
of system control principles, which are challenging and system stability. The experimental analysis demonstrates that the
even infeasible for complex cases. It is time-consuming developed adaptive TSK-type controller outperforms the con-
for complex systems to consider the time-varying and ventional fuzzy logic controllers and the PI controller. A similar
piecewise-linear characteristics, where the controller is TSK-type controller can be found in [31] for approximating the
generally optimized at several critical operational points typical sliding-mode control curve for integrated LED drivers.
rather than the full operational area, resulting in a subop- It is computationally efficient and implemented on a low-cost
timal solution. platform.
2) Once the controller is installed, it operates in a static Although the fuzzy logic controller can handle the system
way with limited adaptability, suggesting that it is only uncertainty, similar to conventional methods such as PID, there
applicable to time-invariant systems. Nevertheless, when is no internal updating mechanism, and thus, the adaptability is
environmental and operational conditions change, the con- limited [50]. Also, it can be seen that the design of member-
troller will be less robust to system parameter shifts and ship functions and fuzzy rules require expert experience, which
the control performance is likely to deteriorate. highly limits the method practicality. Thus, such a method is
3) From the efficient control perspective, an ideal controller applicable to experts only in most cases. Nevertheless, from this
must be able to cope with parameter tolerances with a fast perspective, the expert experience can be coped with fuzzy logic
transient response to maintain system stability. However, and, then, incorporated with other AI techniques as a hybrid
such a desired feature cannot be well fulfilled. method, as discussed later.
These limitations can be mitigated with AI methods. For the
regression-related task in control applications, it is organized in
terms of fuzzy logic, NN, and RL. B. NN-Based Controller
As a black-box technique, NN can approximate a wide range
of nonlinear functions to arbitrary accuracy. With few require-
A. Fuzzy Logic-Based Controller ments on system knowledge, the NN-based controller possesses
Fuzzy logic-based methods have been widely applied to the several advantages, such as robustness, model-free, dynamic,
control of power electronic systems, e.g., speed control [30], adaptive, universal approximation, etc.
MPPT [35], energy management [149], to name a few. 1) Conventional NN: The most widely used NN in power
In [30], a control strategy with three fuzzy logic controllers electronics is the FFNN (or backpropagation NN) with a feed-
is developed for a variable speed wind generation system. The forward multilayer and a backpropagation topology [14]. The
structure of the generator speed programming controller is respective applications essentially exploit the property of static
given in Fig. 8. The control variables include the increment of nonlinear mapping of the FFNN.
the output power ΔPo and the last variation of speed LΔwr∗ . In [82], an FFNN is applied to the waveform processing
The controller outputs the variation of speed Δwr∗ to adjust and delayless filtering. With two cases of variable frequency
the generator speed for a maximum wind power output. The and variable magnitude, it indicates that the FFNN can convert
Mamdani-type fuzzy logic is applied and the information is m-phase waveform with an arbitrary shape into the n-phase
aggregated according to the rule matrix table, e.g., “IF ΔPo waveform with various characteristics of magnitude and fre-
is PS AND LΔwr∗ is ZE, THEN Δwr∗ is PM.” The membership quency. The FFNN-based waveform processing method pro-
functions are iteratively tuned by the system simulation and vides a simplification of the hardware implementation. More-
experiment. Similar Mamdani-type fuzzy logic controller for over, additional single processing functions can be embedded
the primary frequency regulation of a wind farm can be found easily due to the structure flexibility.
in [34]. In [83], the space vector PWM (SVPWM) for a three-level
In [36], a fuzzy logic controller is proposed for regulating voltage-fed inverter is implemented with an FFNN. The input
the speed of a switched reluctance motor based on TSK fuzzy of the NN is the sampled command phase voltages and the
logic by approximating an ideal control law. The parameter output is the pulsewidth patterns of SVPWM. The training
4644 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

Fig. 9. Structure of an RBFN with three layers [50]. x1i is the input of the
input layer node i and yi1 is its output. yj2 is the output of the hidden layer node
j. yk3 is the output of the output layer node k. The input layer and the hidden
layer are fully and directly connected with no weights.

data are generated by the simulation with an SVPWM algo-

rithm. By comparing with a conventional digital signal processor
(DSP)-based SVPWM solution, the performance of the FFNN-
based SVPWM is verified and it can be flexibly implemented on
a dedicated IC chip.
In addition to FFNN, another conventional NN structure is
radial basis function network (RBFN). In FFNN, the weights
of input-to-hidden and hidden-to-output are simultaneously de-
termined. For RBFN, the input layer is directly and fully con-
nected to the hidden layer without weights. The hidden layer is
connected to the output layer by weights Wj , which are the only
weight parameters to be determined in the training, as shown
in Fig. 9. Typically, the generalization of RBFN is better than
FFNN and the training speed and the execution speed are faster.
An exemplary application of RBFN in a three-phase induction Fig. 10. FNN-based controller for a boost converter [100]. x1 is the sliding
generator to regulate the dc-link voltage and the ac line voltage surface S(x) and x2 is its differentiation, n = 2. µji is the jth membership
can be found in [50]. function for input xi . w is the weight between layers. (a) Block diagram of the
FNN-based controller for a boost converter. (b) FNN with a four-layer structure.
Regarding the number of neurons, there are few principles to
determine the optimal number. A generic method is to start with
a relatively small number of neurons and then gradually increase
it according to the training error. For the activation function in the In [100], an FNN is applied to simulate the sliding-mode con-
hidden layer, there are various options, including sigmoid [4], trol of a boost converter to alleviate the chattering phenomena.
[51], [52], [83], radial basis function [50], [150], hyperbolic The block diagram of the controller is given in Fig. 10(a) and the
tangent function [106], [151], wavelet [46], [53], [84], [152], FNN structure with four-layer is given in Fig. 10(b). The inputs
etc. It is worth mentioning that the wavelet activation function of the FNN include the sliding surface S(t) and its differentiation
possesses the superior capabilities of convergence speed and Ṡ(t), which are obtained based on tracking the errors of the
generalization. average output voltage ev and inductor current ei , given the
2) NN With Fuzzy Logic: In control applications, parameter reference voltage Vref and current iref . The output control signal
uncertainty and external disturbance should be well considered is the duty cycle u of PWM.The fuzzy inference is implemented
for system stability and robustness. As a result, an improved by the rule layer as lk = ni=1 wji k j
μi (xi ). The network output
 Ny
variant of NN, i.e., FNN, or neurofuzzy system, which is a is obtained as u = f ( k=1 wk lk ). For the voltage control, the
hybridization of NN and fuzzy logic, is proposed. FNN has the voltage tracking performance is evaluated by the mean-square
merits from both aspects [100], i.e., the humanlike IF–THEN error (MSE) of the output voltage
reasoning rules of fuzzy logic that incorporates expert knowl-
edge and cognitive uncertainty, and the strong capabilities of
1 2
T
approximation and generalization to any nonlinear systems by MSE = ev (d) (3)
the NN. More theoretical details of FNN can be found in [39]. T
d=1
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4645

Fig. 11. ANFIS-based controller for a PWM-inverter-fed induction motor

drive [101]. It is a five-layer network structure with the capability of automatic
identification of fuzzy rules.

where T is the number of sampling instants. The network tuning Fig. 12. RFNN controller for the high-precision trajectory tracking control
aims to reduce the MSE as much as possible to output an of a linear microstepping motor driver [99]. A memory unit of time-delayed
accurate and stable voltage. The performance of the FNN can feedback connection Z −1 is added to enable the dynamic capability of NN
controller.
be significantly improved if the membership function is well
designed. For example, in [46], an asymmetric membership
function is applied to the controller of a six-phase permanent applicable to the static relationship mapping and behavior char-
magnet synchronous motor. It indicates that the learning speed acterization. The dynamic performance of the controller is crit-
can be improved and the network structure can be simplified ical for the transient response. To enable the dynamic capability
compared to conventional membership functions, e.g., Gaussian of the NN controller, a memory unit of time-delayed feedback
function [71], [99], [100]. connection Z −1 is usually inserted to formulate recurrent NN
One of the challenges of FNN is the design of the fuzzy rule, (RNN) [107], as shown in Fig. 12. The outputs of the network not
where extensive expert experience is usually needed [100]. To only depend on the present inputs but also on the previous ones.
overcome this challenge, another typical and effective frame- As a result, the network structure can tackle the time series data
work incorporating fuzzy logic and NN is an ANFIS, which can to facilitate the better performance of dynamics and sensitivity.
be extended from the four-layer structure in Fig. 10 as a five-layer In [106], a robust controller based on RNN is proposed for
topology [101], as shown in Fig. 11. In the ANFIS, the IF–THEN single-phase grid-connected converters for better control perfor-
fuzzy rules, which require the involvement of experts, can be mance in the presence of system parameter changes. The training
generated automatically in the training. For example, in [101], of the RNN is completed by the Levenberg–Marquardt (LM)
a direct-torque neurofuzzy control scheme is developed for a method [13], [82], [106]. The harmonics can be significantly
PWM-inverter-fed induction motor drive based on an ANFIS. reduced by using the proposed RNN-based controller, and the
As shown in Fig. 11, the inputs of the ANFIS-based controller requirements of the high sampling and switching frequency and
include the flux error εm and the torque error εΨ . Layer 1 the damping policies for the conventional control methods can
is the membership layer with the input weights wm and wΨ . be mitigated. A similar RNN structure, which is also termed as
Layer 2 chooses the minimum from the inputs. Normalization Elman NN (ENN), can be found in [52].
is performed in layer 3. In layer 4, the outputs oi is linearly In addition to the performance of dynamics, fuzzy logic is
combined with the network inputs ud = (εm , εΨ ). Layer 5 is also incorporated into RNN in order to improve the performance
the network outputs of the stator voltage command vectors of robustness. For example, in [99], a controller based on a
in polar coordinates Vc and ϕVc . Δγi is the increment angle TSK-type self-organizing recurrent FNN (RFNN) is proposed
and γs is the actual angle of the stator flux vector. In contrast for a high-precision trajectory tracking control of a linear mi-
to the conventional training schemes, the parameter tuning of crostepping motor driver. The network structure is given in
the ANFIS is completed interactively with the backpropagation Fig. 12. The TSK-type self-organizing RFNN is applied to model
algorithms (for membership functions) and the least square the inverse dynamics of the driver. Compared to the FNN in
method (for parameters in fourth layer). More theoretical details Fig. 10(b), the key of the RFNN is the insertion of a recurrent
of the training methods of the ANFIS can be found in [153]. layer, where the delayed neuron output hi (k) is returned as the
3) NN With Recurrent Units: The NN structures in neuron input to facilitate the network dynamics. The network
Section IV-B1 and FNN in Section IV-B2, however, are only diagram and size are adjusted by the self-organizing method, and
4646 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

the respective network parameters are tuned with the method of

recursive least square. As a result, the network diagram and its
parameters can be optimized simultaneously.
4) Training Methods of NN: Essentially, the training of the
NN is an optimization task. Of course, it can be completed
with conventional optimization methods, e.g., PSO [51], re-
cursive least square [99], Kalman filter [105], etc. Considering
a large number of parameters in the NN, these conventional
optimization methods are generally inefficient. As a result, an
elaborate training scheme is developed, i.e., backpropagation Fig. 13. Framework of RL in the MPPT controller of wind energy conversion
algorithm [4], [50], [52], [53], [71], [83], [84], [150]. More systems [5], [138]. A Q-table is formulated to save the optimal generator rotor
theoretical details of the backpropagation algorithm can be speed wr∗ to be performed given the current system state st , including the current
electrical output power Pe and the generator rotor speed wr .
found in [1, Ch. 5].
The backpropagation algorithm is based on the idea of steepest
gradient descent. One of the key steps in the backpropagation
algorithm is the iteration of the weight update of the key steps for the sequential learning is determining a
wk+1 = wk − ηk gk (4) suitable learning rate ηk in (4), since a larger ηk will result in
system instability and a smaller ηk will lead to slow convergence.
where wk is the current weight, gk is the current gradient, ηk is The optimal learning rate ηk can be determined by using the
the learning rate, and wk+1 is the weight of the next iteration. metaheuristic methods in the training, e.g., PSO in [50], [52],
To calculate the gradient gk and find the steepest direction and [53] and differential evolutionary (DE) in [46]. As a result,
of gradient descent efficiently, various improved variants of the sequential learning process can be stable and converges fast.
the backpropagation algorithm have been proposed, e.g., LM
method [13], [82], [106], resilient backpropagation algorithm,
conjugate gradient algorithm, one-step secant algorithm, etc.
Note that it is challenging to determine the most suitable training C. RL-Based Controller
algorithm for a specific task. It depends on multiple factors, With RL, the controller learns a goal-oriented control strat-
including problem complexity, dataset size, number of param- egy by interacting with the physical system or its simulation
eters, task types of classification or regression, etc. A useful model [138]. It accumulates experience progressively and learns
reference can be found in MATLAB Manual of Neural Net- a specific control strategy that maximizes predefined goals.
work Toolbox [40], where the theoretical details, advantages, One of the relevant applications of RL-based controller is the
limitations, and comparisons of these training algorithms are MPPT in renewable energy systems [5], as shown in Fig. 13.
thoroughly analyzed with several benchmark examples. It is Specifically, a real-time intelligent MPPT algorithm using RL
worth mentioning that the LM method is one of the most widely is proposed for a wind energy conversion system. With the
used methods for the applications in power electronics with a online learning capability of RL by interacting with the environ-
fast convergence speed and a high accuracy. ment, an optimum control strategy is formulated in the Q-table.
Considering whether the training dataset is available in a batch The Q-table consists of elements of state transition probability
form or in a sequential form, the training scheme of the NN can q(st , at ), which can facilitate the maximized power output (or
be completed in either batch learning, which is also termed as reward) if action at , i.e., the expected generator rotor speed
offline learning, or sequential learning, which is also termed as wr∗ , is performed given the current system state st , including
online learning or incremental learning. the current electrical output power Pe and the generator rotor
For batch learning, the gradient gk in (4) is calculated based speed wr . As a highlight, the wind turbine parameter and the
on all the data points in the dataset for the parameter updates. It wind speed are not required. This article is further extended
generally applies to the case where the whole dataset is available by integrating an NN into the Q-learning of RL [6]. In this
before the NN is implemented for field application, e.g., the way, the challenges in the determination of the state space are
waveform processing and delayless filtering in [82]. avoided. The online learning process can be reactivated once the
For sequential learning, the gradient gk in (4) is calculated learned optimal relationship is destructed by the system aging
based on every newly available data point or several newly avail- behaviors. It significantly improves the autonomous capability
able data points forming a minibatch. Therefore, the learning of the wind energy conversion system. A similar example can
process is incrementally completed. This feature is especially be found in [140], where RL is applied to the MPPT control of
useful for the case where the training data can only be sequen- a buck converter of PV arrays.
tially obtained in field applications. The intelligent controller For the NN-based controller, the learning process is completed
[53] is a typical case of a sequential training scheme since the from examples provided by an external supervisor. While the
input data of the NN can only be available sequentially by inter- RL controller can learn the experience by directly interacting
acting with the output of the control command and the system. with the environment through actions and rewards. It is worth
With this adaptive capability, the NN can be reparameterized mentioning that the training of the RL controller is based on
and reconfigured for tracking the system parameter shifts. One the interactions between the controller and the system, and the
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4647

Fig. 14. Flowchart of maintenance in power electronic systems.

TABLE IV complex and severe working environments [18], [154], [155].

ADVANTAGES AND LIMITATIONS OF AI ALGORITHMS IN
CONTROL APPLICATIONS
The reliability and safety of power electronic components, con-
verters, and systems are of great importance for field applica-
tions. In maintenance, preventive activities, including condition
monitoring, anomaly detection, fault diagnosis, RUL prediction,
etc., are effective approaches to ensure that intended functions
can be properly executed. These activities are aligned with the
IEEE standard framework of PHM for electronic systems [156].
Fig. 14 presents a systematic flowchart of maintenance activities
in power electronic systems. Generally, it consists of following
three parts.
1) Offline Training And Knowledge Learning: It integrates
various aspects of knowledge including historical moni-
toring data, simulation data, accelerated aging test experi-
FFNN—Feed-forward NN and its variants, FNN—Fuzzy NN and its variants, RNN— ment, failure mode and effects analysis, etc. Moreover, en-
Recurrent NN and its variants, RFNN—Recurrent fuzzy NN and its variants. Superior: semble methods or fusion techniques are typically applied
+++, intermediate: ++, inferior: +. to this part for performance improvement. As a result,
physical system dynamics and behaviors (e.g., degrada-
offline dataset is unnecessary in this case. As a result, the RL- tion behavior) can be accurately characterized as offline
based controller is beneficial to new systems without existing models based on the information of the unit population.
datasets. 2) Condition Monitoring and Health Assessment: This part
deals with the health assessment of the unit in service
D. Discussions subjected to the online condition monitoring in field ap-
plications. The offline model is tailored and individualized
A summary of the advantages and limitations of AI algorithms
to the unit in service through the model parameter tuning
in control applications is given in Table IV. It is worth mention-
layer by adapting to field operational environment and
ing that the dynamic performance, robustness, generalization,
workload. The functions include the noninvasive parame-
and convergence speed of AI algorithms are critical in con-
ter identification, data preprocessing (e.g., data cleaning),
trol applications. The algorithm complexity and computational
feature mining, anomaly detection, fault diagnosis, and
burden are the major challenges. Thus, high-performance DSP
RUL prediction. In this way, insightful knowledge for
or field programmable gate array is necessary for practical
decision-making can be extracted from the continuous
implementations.
condition monitoring information.
3) Management and Decision-Making: In this part, the sup-
V. MAINTENANCE portive knowledge of health assessment is returned for
Although reliability characteristics have been elaborately con- optimal decision making. With this feedback, control
sidered in design and control, power electronic systems still policies (e.g., power routing) can be adjusted to maxi-
undertake various risks and even catastrophic failures due to mize the system performance given the real-time health
4648 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

status. Moreover, economical maintenance policy can

be made to facilitate the condition-based and predictive
maintenance.
Subsequently, the relevant applications of AI in maintenance
in terms of these three parts are discussed in detail.

A. Condition Monitoring
Condition monitoring [20], [157], [158] in power electronics
includes system parameter identification, data preprocessing,
and feature mining. The condition monitoring information is
applied to uncover hidden and informative insights, which serve
as a basis for the subsequent PHM applications.
1) System Parameter Identification: The system parameter Fig. 15. Examples of model-free methods of system parameter identification
identification [159] deals with information acquisition for crit- with AI. (a) Capacitance identification of dc-link capacitor [88]. (b) a-phase cur-
rent estimation for calculating the impedance measurement of power electronic
ical components. Developing specific hardware for parameter system [108]. (c) ESR estimation in future p steps for supercapacitors [103].
identification (e.g., temperature-sensitive electrical parameters
of IGBTs [158]), however, is quite a challenging task due to
features of power electronic systems, e.g., very tight space in a include the supply voltage Vt , the supercapacitor temperature
power module, very fast switching frequency, relatively insignif- θt , and a time series ESRt−400:100:t consisting of five previous
icant parameter changes in terms of aging [157], etc. One of the ESR data points. The output of the ANFIS is the ESR estimations
promising solutions is noninvasive method without any extra in future p steps. Experimental analysis indicates that ESR of
hardware implementation, where information of interest can be supercapacitor can be accurately estimated and the normalized
inferred or estimated indirectly from available physical signals. root-mean-square error of the ESR estimation is as small as
As a result, the condition monitoring can be implemented with 0.025 at condition monitoring time of 2600 h.
a sensorless and cost-efficient solution, which is favorable for A summary for the framework of model-free parameter iden-
industrial practitioners. Generally, the system parameter iden- tification methods is given in Fig. 15. It can be seen that AI
tification can be categorized into model-free and model-based methods serve as the regression tool f (·) between the available
methods, considering whether the system dynamics and models input signals and the parameter to be monitored.
are required. The model-free method is attractive for industrial applications
For the model-free method, no prior knowledge of the system due to less hardware cost. However, it is typically sensitive to
dynamics is required. Essentially, it deals with the regression external noise and disturbance due to the lack of system model.
capability of AI algorithms to construct a relationship between Thus, its robustness should be carefully considered. This issue
the inputs and outputs. For example, in a three-phase front-end can be possibly mitigated with a large amount of data in the train-
diode bridge motor drive, the current ia,out in a-phase and the ing stage [159] to cover situations in field applications as much
dc-link ripple voltage Δvdc are considered as the inputs, and as possible. Nevertheless, the data collection is time-consuming
the capacitance C is applied as the output for the training of an and costly.
FFNN [86]–[88]. In this way, the relationship between the input Another category of the system parameter identification is the
signals and the capacitance is established and, thus, the capaci- model-based method. As the name implies, for a model-based
tance can be inferred indirectly. Similarly, it is demonstrated that method, system physics and models are partially known in ad-
the capacitance can be estimated by the FFNN constructed by vance and the identification model is formulated with unknown
the frequency domain information of dc-link voltage ripple. The model parameters. In this way, the system identification task
potentials of FFNN in the capacitance estimation are illustrated is equivalent to the exploration of optimal parameters of the
in a hardware prototype [88]. model, which is essentially an optimization task. In this case, AI,
In [108], considering the dynamic capability of RNN, an especially the metaheuristic methods, is utilized as an optimizer
impedance identification method is proposed based on RNN to to find the optimal solutions. Numerous approaches such as
enable the stability analysis for power electronic systems over a PSO [57], crow search algorithm [73], GA [69], etc., or their
wide frequency range. The RNN is applied to build a model that improved variants, can be exploited.
can produce identical outputs as the physical system given the In [69], a parameter identification method for the health
same inputs. The inputs of RNN include three-phase voltages diagnostic of a PV panel is developed. The equivalent circuit
va , vb , vc . The output is the a-phase current ia . As a result, the of the PV panel is given in Fig. 16, and its system model is
RNN-based model possesses the same frequency characteristics explicitly derived as
as the physical one. It can be performed for the impedance ⎧ vsh vsh − v
⎪
⎨iCsh (vsh , v) = Iph − iD (vsh ) − − ,
identification without interrupting the system operation. Rsh Rs

In [103], an improved ANFIS is applied to estimate the ⎪ dvsh (vsh , v) 1 vsh vsh − v
⎩ = Iph − iD (vsh ) − −
capacitance and equivalent series resistance (ESR) of the su- dt Csh Rsh Rs
percapacitor. At monitoring time t, the inputs of the ANFIS (5)
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4649

Fig. 16. Dynamic model of a PV panel for parameter identification with the
model-based method [69]. System parameters include the input current Iph ,
output current Io (i(t) ), voltage vsh across capacitor Csh , resistor Rsh , p-n
junction capacitance Csh , and resistor Rs .

where Iph is the input current, Io is the output current, vsh is

the voltage across the capacitor Csh , Rsh is the resistance, and
Csh is the p-n junction capacitance. As a result, the param-
eter identification is equivalent to find a parameter set G = Fig. 17. k-means clustering method for discretization of filtered degradation
{Iph , Io , vsh , Rsh , Csh , Rs } that ensures an identical output as paths of increment of drain-to-source ON-state resistance ΔRDS(on) of different
the physical system. By injecting large signal disturbances to power MOSFET devices #26, #29, #32, #35, #36, #37, #38 [131].
the panel voltages in the testing stage, the dynamic response of
the current–voltage characteristics is sampled to calculate the fusion to integrate multiple information sources, etc. Typically,
objective function as the performance of the subsequent PHM application, e.g., the

N2 diagnostic accuracy, can be significantly improved if the data
1
fobj (G) = (ip [k] − i[k])2 (6) preprocessing and feature mining are properly conducted.
N1 − N 2 + 1 In [131], a reliability assessment method for power MOSFETs
k=N1
based on a continuous-time Markov chain is proposed. To
where ip [k] and i[k] are the current output of the model and the
discretize the continuous degradation path of power MOSFETs
physical system, respectively, and N1 and N2 are the start index
without breaking the inherent monotonicity, a k-means method
and the end index for the sampling, respectively. Subsequently,
is applied to divide the evolution of drain-to-source ON-state
an improved GA method is used to explore an optimal solution
resistance RDS(ON) into 11 discrete states, as shown in Fig. 17.
minimizing fobj (G) in (6). A similar investigation can be found
In [133], a health state identification method for IGBTs based
in [57], where a modified PSO algorithm is applied to the internal
on self-organizing maps (SOMs) is proposed. It is essentially
parameter identification of a PV panel.
a clustering task. The states of the device are clustered as the
Due to the involvement of system dynamics and models, the
healthy state, partially degraded state, heavily degraded state,
amount of data required for the estimation can be significantly
and failure state, considering the distance between the input
reduced for the model-based methods. Also, the overfitting risk
measurements (including collector current Ic , collector–emitter
in the model-free methods can be mitigated. It exhibits better
voltage Vce , and case temperature T ) and the best matching unit
dynamics to handle unexpected disturbance and switchable
of the trained SOMs.
working modes. However, due to the system complexity, the
In [160], a composite failure precursor of SiC MOSFETs is
system dynamics and models are challenging to formulate in
developed with a data fusion technique of genetic programming,
most cases.
which is a variant of GA. It integrates multiple degradation
For parameter identification methods in power electronics, the
signals of a power semiconductor device in a nonlinear way.
accuracy and robustness under the complex environment should
Since the composite failure precursor is directly optimized in
be considered. For example, for the condition monitoring of
terms of the RUL prediction model, the prediction accuracy is
power MOSFETs in [131], the device is considered as failed if
improved by 35.3% and the prediction uncertainty is reduced by
there is an increase of 0.08 Ω for the degradation indicator of
16.3%. It indicates that data fusion in condition monitoring is
drain-to-source ON-state resistance RDS(on) . Such a tiny incre-
potentially useful especially for system-level applications (e.g.,
ment is challenging to observe. Thus, more research efforts are
converters), where multiple physical degradation signals exist.
necessary to improve the sensitivity of the AI-based parameter
An integrated toolbox “Diagnostic Feature Designer” for the
identification methods. Moreover, it is worth mentioning that
feature identification is available in MATLAB [161], which can
computational burden and embedded capabilities should be
be applied to the data preprocessing and feature mining as an
considered for field applications.
automatic tool.
2) Data Preprocessing and Feature Mining: Data prepro-
cessing and feature mining are concerned with refining the
B. Anomaly Detection and Fault Diagnosis
raw data to better serve the applications, e.g., fault diagno-
sis. By exploring dataset structure, it includes data cleaning The anomaly detection makes a binary decision and focuses
to reduce noise, data clustering to discover groups of similar on the abnormal behavior identification. It provides an indication
data points, density estimation to identify the data distribution, when the rated system characteristics or nominal parameters
data compression that projects high-dimensional data down to exceed the predefined safety range. Once the anomaly behavior
low-dimensional data to reduce the number of features, data occurs, the fault diagnosis [19] identifies and locates the detailed
4650 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

RBFN. It indicates that the ESN is superior in the sensitivity,

design process, and training speed.
In [115], an 1-D CNN is applied to the fault diagnosis of a
modular multilevel converter. One advantage of 1-D CNN is
that the feature extraction and diagnostic classification can be
integrated together, which enables the fault diagnostics on the
raw data directly. In this way, the feature extraction, which is
usually experience-intensive, can be avoided. The experimental
Fig. 18. FFNN for anomaly detection of a full-bridge diode rectifier. The results indicate that the proposed method is highly reliable and
inputs of the NN include input voltage vi (t), input current ii (t), and output
current io (t); the output of the NN is output voltage vo (t) [93].
provides a detection accuracy of 98.9% and a fault diagnostic
accuracy of 99.7% within 100 ms.
In addition to the previous NN-based methods, kernel meth-
ods, including the SVM and the RVM, are also applied for
failure modes subsequently. Essentially, anomaly detection and anomaly detection and fault diagnosis. One advantage of the
fault diagnosis are the classification, regression, or clustering kernel methods is that the dataset size requirement is relatively
tasks. Based on the learned relationship from the training stage, lower than the NN-based methods.
it determines the fault label when a new fault signature becomes In [7], based on the time-domain fault features, an SVM-based
available. Note that the feasibility of AI-based anomaly detec- fault diagnosis method is proposed for incipient yet progressive
tion and fault diagnosis is based on two assumptions [33]: First, faults of IGBTs in an inverter. The training of SVM can be
the fault occurrence in any components has an impact on the completed by metaheuristic methods (e.g., PSO, GA, etc.). For
fault signature; second, the impact on these signatures varies a total of 41 fault classes, it achieves an average accuracy of
with different fault modes and fault locations. The methods of 94.82% being robust to both load variations and motor parameter
anomaly detection and fault diagnosis can be categorized as shifts.
supervised learning methods and unsupervised methods. In [127], an RVM is applied for the fault diagnosis of a
1) Supervised Learning Methods: In [93], an FFNN is ap- cascaded H-bridge multilevel inverter. PCA is applied to extract
plied to establish the nonlinear relationship of the inputs and the fault signal feature. Experimental analysis indicates that
outputs of a full-bridge diode rectifier. The training of the FFNN the RVM outperforms the FFNN and the SVM, with 100%
is completed at the normal operation mode of the rectifier, diagnostic accuracy in this specific case study. Compared to
as shown in Fig. 18. As a result, the principles and mapping SVM with the direct fault label as its output, RVM is formulated
relationship between the inputs, including input voltage vi (t), under the Bayesian framework. It makes probabilistic outputs
input current ii (t), and output current io (t), and the output of the fault information, which possesses good theoretical guid-
signal of output voltage vo (t) are characterized, considered as ance and is favorable to the uncertainty analysis on diagnostic
a digital emulator indicating the normal operational mode of results. Generally, for the same task, the RVM is sparser than
the rectifier. This digital emulator and the physical rectifier are SVM, indicating faster speed for field applications. However,
simultaneously operated and their outputs are compared in real the training time of RVM is generally longer than SVM.
time. Once the monitored output voltage of physical rectifier 2) Unsupervised Learning Methods: In [136], PCA is ap-
significantly deviates from the output of FFNN, it suggests that plied to the anomaly detection of SiC MOSFETs. Multiple statis-
the rectifier runs into an abnormal mode, which facilitates the tical features, including kurtosis, skewness, etc., are considered
anomaly detection. In this case, the FFNN essentially serves as as the inputs of the PCA algorithm. The output is compact with
the regression tool. fewer features and a transformation matrix. For field applica-
In [90], an open-circuit fault diagnosis algorithm is proposed tions, the newly available data are applied to the transformation
for the inverter in a microgrid system subjected to varying load matrix for the calculation of an anomaly index. Abnormal be-
conditions. A signal processing method is proposed to reduce the havior is notified when the anomaly index exceeds a predefined
required amount of information for the fault representation and threshold. The method is verified by a processor-in-the-loop ex-
suppress the impact of the load change. An FFNN is used as a periment. This detection mechanism is similar to the that in [93].
diagnostic classifier. The computational burden of the proposed Other unsupervised learning methods in anomaly detection and
method can be reduced to 10% of that of the existing algorithms. fault diagnosis, including k-means and SOMs, can be found
In this case, the FFNN serves as a classification tool. Similar fault in [118].
diagnosis ideas include the ANFIS to determine the severity 3) Discussions: Table V summarizes the features of typical
levels of a capacitor in the dc-link filter [102]. AI algorithms and their variants for anomaly detection and
In [112], a multiswitches fault diagnosis algorithm for fault diagnosis. It can be seen that each AI algorithm possesses
voltage-source inverters is proposed. An echo state network advantages and limitations. To fully exploit the advantages of
(ESN) is used as a diagnostic classifier given small low- each algorithm, it is effective to combine multiple algorithms
frequency data. Note that ESN is an improved variant of RNN for a decision-level fusion to improve the diagnostic accuracy
to avoid gradient exploding and vanishing in the training. In this and robustness. An example of decision-level fusion for fault
work, the diagnostic performance of ESN is compared with the diagnosis of IGBTs can be found in [96]. More ensemble
FFNN, the FFNN with a wavelet activation function, and the methods to combine multiple algorithms can be found in
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4651

TABLE V
ALGORITHM COMPARISONS OF ANOMALY DETECTION AND FAULT DIAGNOSIS IN POWER ELECTRONIC SYSTEMS

N/A: not available.

[1, Ch. 14]. From the AI perspective, there is a negligible dif-

ference between power electronics and other engineering areas
(e.g., electromechanical applications) in terms of the anomaly
detection and fault diagnosis tasks. Two reviews of AI methods
in anomaly detection and fault diagnosis can be found in [162]
and [163].
Note that various AI methods and their variants have been
successfully applied to anomaly detection and fault diagnosis.
There are differences in terms of how the data are collected
and types of available data in different applications, which is
an important aspect of practical applications of AI. An in-
tegrated platform “Predictive Maintenance Toolbox” is avail-
able in MATLAB [164], which includes various algorithms of
anomaly detection and diagnostics. It is beneficial for the method
development and benchmark analysis. From the AI perspective, Fig. 19. Flowchart and procedures of AI methods for RUL prediction of power
most of the methods can be interchangeably applied with a electronic systems, where pdf: probability density function.
comparable performance in terms of the evaluation accuracy.
Although the accuracy can be further improved by advanced
algorithms (e.g., deep learning methods), the accuracy improve- C. RUL Prediction
ment after a high score, e.g., 90%, is relatively less significant Lifetime prediction in the design phase is to support the DfR,
compared with other practical concerns. More considerations which refers to the feature of a population of units. As one
should be devoted to the gap between theoretical algorithms and of the critical aspects of PHM [165], the RUL prediction is
practical implementations, where the practical considerations not to predict the lifetime of a population of units. It predicts
include the following. the residual lifetime of an individual unit in service based
1) In addition to the single component fault, the failure mode on the condition monitoring information. There are associated
of multiple components failed simultaneously should be uncertainties in the lifetime prediction, including model calibra-
considered. The dependence and coupling effects among tion errors, manufacturing tolerances, variations of operational
the component failures should be incorporated into the environments and workload, etc. These uncertainties result in
diagnostic algorithms. inaccurate reliability estimates for a specific unit in field oper-
2) Considering the challenges in the data acquisition of power ation [166]. RUL prediction is applied as an additional tool to
electronic systems, the training dataset for practical appli- reduce the uncertainties for reliability-critical, safety-critical, or
cation is typically limited. This situation is even worse availability-critical applications.
for a dataset with unbalanced fault labels, i.e., the ample The flowchart and procedures for RUL prediction are given
data of the normal operation case and the scarcity of data in Fig. 19. The regression model can be established based
with fault labels due to catastrophic failures. Thus, the on historical dataset. The probability density function (pdf) of
algorithm applicability given limited size of dataset and degradation level at any specific condition monitoring time can
poor quality dataset should be investigated. be estimated based on the regression model. The pdf of the
3) The practicality, including computational burden, adaptive RUL can be derived from the pdf of the degradation level.
capability, robustness, difficulty of algorithm design and Given the fact that the system is properly functioning at con-
debugging [112], implementation cost, etc., should also dition monitoring time t, its RUL l is defined as the residual
be comprehensively considered. lifetime when the degradation process D(t) exceeds the failure
4652 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

Fig. 20. RUL prediction of power MOSFETs based on ESN [111]. For the
network training, the input weights W in and the recurrent weights W are
randomly generated. The output weights W out are estimated by least-square
methods. Fig. 21. Gaussian processes regression for the RUL prediction of IGBTs [119].

threshold w, i.e.,

l = inf {l : D(t + l) ≥ w | D(t) < w, D1:j } (7) uncertainty quantification is more critical for RUL predic-
tion. As shown in Fig. 19, the RUL is a random variable
where D1:j is the cumulative CM information up to time t. Note and, thus, quantification of the confidence interval is essen-
that RUL l is a random variable. In addition to its expected value, tial for the optimal decision-making. These uncertainties
the uncertainty metrics with the lower and upper confidence come from the population heterogeneity, measurement
interval (llo , lup ) are also of great importance. AI methods in noise, varying operational settings, etc., which should be
RUL prediction is typically dealing with a nonlinear regression comprehensively considered for a practical solution. AI
between the degradation information and the corresponding methods are rather challenging for the uncertainty quan-
RUL based on the training dataset [167]. In this way, degradation tification of prediction results considering the black-box
patterns can be characterized. Once the degradation patterns feature. Several feasible approaches include the Monte
have been learned, it can be directly projected based on the Carlo methods [114], incorporating particle filter in the
regression model to facilitate the future degradation level pre- NN [111], and Bayesian-based AI methods (e.g., Gaus-
diction. As a result, the RUL can be estimated. sian process, RVM). Another promising direction is the
In [111], an ESN is applied to the RUL prediction of power stochastic data-driven methods [154], [160], [168], which
MOSFETs, as shown in Fig. 20. The input of the ESN is the degra- can intrinsically provide the pdf of the RUL for calculating
dation indicator drain-to-source ON-state resistance RDS,(on) at the confidence interval.
times k − 1 and k, and the output is the RDS,(on) at time k + 1. To 2) Adaptive capability: It is concerned with the model pa-
facilitate the adaptation of the ESN, a particle filter is exploited rameter tuning layer in Fig. 14 for connecting the offline
to recursively update the output weights when new condition models and the online models, which is a key step for
monitoring data of the in-situ device becomes available. In this practical applications. If a specific AI method lacks an
way, the degradation model is adaptive to varying external envi- adaptive capability, its application is limited since one
ronments and operational modes. Another NN method involving prerequisite is that the training data and the test data should
TDNN for the RUL prediction of IGBTs can be found in [114]. be generated under similar situations (e.g., external envi-
In [119], Gaussian processes regression is applied to the RUL ronments and operational modes) and share a high-level
prediction of IGBTs. For the degradation modeling, the nonlin- similarity [95]. It is challenging for power electronics
ear relationship between the decrement of ON-state collector– since operational settings of the in-situ system (i.e., the test
emitter voltage ΔVce,on and the condition monitoring time is data) are quite different from that of the training dataset,
established by the Gaussian processes regression. Since Gaus- which is generally obtained with accelerated testing ex-
sian process is formulated with the Bayesian framework, it is periments. The majority of the research [74], [114], [119]
able to predict the uncertainty of variation ΔVce,on intrinsically. assumes that the operational settings of the in-situ system
It can be seen from Fig. 21 that the error bar of the evolution of are identical to the training dataset (e.g., accelerated aging
ΔVce,on is explicitly derived, which can be further utilized for the experiments), which may not be the case in field applica-
calculation of the confidence interval of RUL. Another example tions. Thus, the adaptive capability of the AI-based RUL
of the kernel method for RUL prediction can be found in [74], prediction method is critical to bridge academic research
where an SVM is applied to the degradation modeling of a buck and industrial applications. Other promising directions
converter. of model parameter tuning include the explicit mapping
To make AI-based methods of the RUL prediction more relationship derivations [169] and transfer learning [170],
practical for field applications, more efforts should be devoted [171] of degradation characteristics under various opera-
to the following aspects. tional settings (temperature, voltage, humidity, etc.). This
1) Uncertainty quantification: Compared to other regression- may, however, imply intensive investigations of system
related tasks, e.g., control applications, the capability of models.
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4653

TABLE VI 1) Motivations and Justifications of AI Applied to Power

REQUIREMENTS OF AI FOR EXEMPLARY APPLICATIONS IN DESIGN,
CONTROL, AND MAINTENANCE
Electronic Systems: Although there are numerous stud-
ies on AI for power electronic systems in the literature
since the 1990s, the practical implementations in indus-
try are still limited, which is a sharp contrast compared
to the claimed AI potentials. It is necessary for deeper
investigations into tasks where AI can essentially outper-
form conventional methods. The justifications of AI-based
solutions should be clearly identified by comparing to
conventional methods from the industrial perspectives,
High: +++, moderate: ++, low: +.
e.g., implementation complexity, algorithm accuracy and
robustness, algorithm accountability, extra hardware cost,
computational energy consumption, embedded capability,
etc.
VI. OUTLOOK ON AI FOR POWER ELECTRONIC SYSTEMS 2) Interwoven AI Implementations Through Life-Cycle
From the algorithm perspective, it is necessary to investigate Phases: Implementations of AI in each life-cycle phase
the features of AI when it applies to different life-cycle phases. of design, control, and maintenance will facilitate flexible
By using a power converter system, three specific examples are functional interactions. This feature is beneficial to overall
applied to illustrate the requirements of AI methods for each performance optimization and procedure simplification.
life-cycle phase. It enables the system capability in managing data flow
For the heatsink design of a converter system, a large num- between electrical and other disciplines (e.g., mechanical
ber of decision variables, e.g., weight, volume, pattern, need area) [13] as well. For example, aging information ob-
to be determined, which is essentially an optimization task. tained by the AI-based system parameter identification can
The metaheuristic methods are applied to the optimization that be flexibly incorporated into the AI-based controller for
involves an iterative trial-and-error procedure. Although the reliability improvement. Therefore, more attention should
computational effort is intensive, the design task is typically be paid to the interwoven interactions powered by AI.
performed offline. There is less requirement on the algorithm 3) Multilevel Information Fusion: Robustness is essential for
speed in this case. Although the metaheuristic method-based safety-critical power electronic systems. Multiple sources
optimization does not ensure a global solution, the suboptimal of information and models are available in most cases
heatsink design is still superior and satisfactory in most cases. for a specific application of power electronic systems. If
Thus, the algorithm accuracy is not critical as well. The training these information sources and models are simultaneously
dataset and interpretability of the optimization process are not exploited, possible biases can be mitigated to improve ro-
required. bustness. Multilevel information fusion can be performed
For the intelligent controller of a converter system, the real- at the data-level [160], [172], feature-level, decision-
time control errors, e.g., the voltage error, the current error, need level [96], and their combinations, in order to exploit the
to be returned to the controller for the adaptive updating in an insights of each information sources. For example, the
online mode. Thus, the requirements of algorithm speed and well-established differential equations of power converter
accuracy are the most critical. In addition, the controller stability system can be integrated with AI as a hybrid solution for
needs to be theoretically ensured, and thus, the interpretability condition monitoring. As a result, the advantages from
is critical. The intelligent controller is generally tuned online, it both the model-driven side and the data-driven side can
is unnecessary to prepare the dataset for the model training. be gained for better accuracy and robustness.
For the RUL prediction of switching devices in a converter 4) Computation-Light AI: Compared to other industrial areas
system, the requirement of the algorithm speed is moderate since (e.g., image recognition), one of the key features of power
the device degradation is slow and the long time span of decision electronic systems is that there is no powerful computation
making is acceptable. The degradation model for the RUL unit. While real-time applications, e.g., control, impose
prediction can be prepared in offline mode and efficiently tuned a rigid requirement on the algorithm speed. Although
in online mode, and the computational effort in this application complex deep learning techniques [170] can provide su-
is moderate. Since the model accuracy is highly dependent on perior performance, it is computationally intensive for
the dataset, the dataset requirement, e.g., dataset quality, dataset power electronic systems. A prospective direction is the
size, label balance (e.g., limited abnormal data in the training computation-light AI algorithms that can be implemented
dataset), etc., is the most critical. Moreover, the interpretability on cost-effective units but provide comparable perfor-
of the RUL prediction results with uncertainty is critical as well. mance with deep learning algorithms.
As a result, a comparison of AI algorithms in each phase of the 5) Data-Light AI: One of the bottlenecks of AI implemen-
life-cycle of power electronic systems is provided in Table VI. tation on power electronic systems is the dataset. For
It is concluded that AI possesses immense potentialities in example, AI-based solutions for RUL prediction require
power electronic systems. Many opportunities and issues are the dataset to be versatile enough for accurate degradation
yet to be explored as follows. behavior learning. However, the dataset size is generally
4654 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

small since the degradation experiments are resource- 3) From the function perspective, the AI-related applications
consuming. This situation is even worse for safety-critical are essentially dealing with the optimization, classifica-
cases. Thus, developing AI algorithms with lower dataset tion, regression, and data structure exploration.
requirement, i.e., data-light AI solutions that can provide 4) The milestones of relevant algorithm variants and appli-
acceptable performance in the presence of poor datasets, cations are identified and organized as a timeline map.
is a prospective direction. 5) For each life-cycle phase, illustrative examples are dis-
6) Explainable AI: Most of the AI algorithms in power cussed and the challenges and future research opportuni-
electronics suffer from the “black-box” feature. For ex- ties are identified.
ample, most of the AI-based solutions for RUL predic-
tion can only provide a point estimation without sensi-
REFERENCES
tivity analysis and uncertainty quantification. It makes
AI-based solutions opaque and less convincing for prac- [1] C. M. Bishop, Pattern Recognition and Machine Learning. Berlin, Ger-
many: Springer-Verlag, 2006.
titioners to implement in industry applications, especially [2] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge,
for safety-critical cases. There is a pressing need to im- MA, USA: MIT Press, 2016.
prove the algorithm transparency for explainable AI with [3] T. Wu, Z. Wang, B. Ozpineci, M. Chinthavali, and S. Campbell, “Auto-
mated heatsink optimization for air-cooled power semiconductor mod-
better interpretability. Understanding how models come ules,” IEEE Trans. Power Electron., vol. 34, no. 6, pp. 5027–5031,
up the decisions is critical for model simplification and Jun. 2019.
safety, with which AI solutions can be implemented with [4] X. Zhan, W. Wang, and H. Chung, “A neural-network-based color control
method for multi-color LED systems,” IEEE Trans. Power Electron.,
confidence. vol. 34, no. 8, pp. 7900–7913, Aug. 2019.
7) Dataset Privacy: An increasing attention has been paid [5] C. Wei, Z. Zhang, W. Qiao, and L. Qu, “Reinforcement-learning-based in-
to the data privacy, e.g., General Data Protection Regu- telligent maximum power point tracking control for wind energy conver-
sion systems,” IEEE Trans. Ind. Electron., vol. 62, no. 10, pp. 6360–6370,
lation [173] in the European Union. With these critical Oct. 2015.
regulations, the training of standard AI algorithms is [6] C. Wei, Z. Zhang, W. Qiao, and L. Y. Qu, “An adaptive network-based
challenging since a centralized data collection may be reinforcement learning method for MPPT control of PMSG wind energy
conversion systems,” IEEE Trans. Power Electron., vol. 31, no. 11,
not feasible in the future. Thus, for power electronics pp. 7837–7848, Nov. 2016.
applications, it is promising to develop a collaborative [7] I. Bandyopadhyay, P. Purkait, and C. Koley, “Performance of a clas-
learning scheme for AI algorithms without collectively sifier based on time-domain features for incipient fault detection in
inverter drives,” IEEE Trans. Ind. Informat., vol. 15, no. 1, pp. 3–14,
aggregating data from different locations, e.g., federated Jan. 2019.
learning [174]. It is well aligned with the trend of data [8] A. E. Mejdoubi, H. Chaoui, J. Sabor, and H. Gualous, “Remaining useful
privacy regulations for the implementation of AI solutions. life prognosis of supercapacitors under temperature and voltage aging
conditions,” IEEE Trans. Ind. Electron., vol. 65, no. 5, pp. 4357–4367,
8) Power Electronics Database: Due to the complexity of May 2018.
system dynamics of power electronics, extensive datasets [9] F. Tao, H. Zhan, A. Liu, and A. Y. C. Nee, “Digital twin in industry: State-
are required for the model training, especially for the of-the-art,” IEEE Trans. Ind. Informat., vol. 15, no. 4, pp. 2405–2415,
Apr. 2019.
maintenance applications. While the experimental testing [10] X. He, W. Shi, W. Li, H. Luo, and R. Zhao, “Reliability enhancement of
for data collection is generally time-consuming and expen- power electronics systems by big data science,” in Proc. Chin. Soc. Elect.
sive. There is a compelling demand for building up com- Electron. Eng., vol. 37, no. 1, pp. 209–221, Jan. 2017.
[11] K. L. Tsui, Y. Zhao, and D. Wang, “Big data opportunities: System health
mon power electronics data and knowledge base. These monitoring and management,” IEEE Access, vol. 7, pp. 68853–68867,
open-source datasets are critical to benchmark algorithm May 2019.
performance and accelerating application development. It [12] S. E. De Leon-Aldaco, H. Calleja, and J. A. Alquicira, “Metaheuristic
optimization methods applied to power converters: A review,” IEEE
will benefit the global power electronics communities in Trans. Power Electron., vol. 30, no. 12, pp. 6791–6803, Dec. 2015.
academia and industry. [13] M. R. G. Meireles, P. E. M. Almeida, and M. G. Simoes, “A compre-
hensive review for industrial applicability of artificial neural networks,”
IEEE Trans. Ind. Electron., vol. 50, no. 3, pp. 585–601, Jun. 2003.
VII. CONCLUSION [14] B. K. Bose, “Neural network applications in power electronics and motor
drives-an introduction and perspective,” IEEE Trans. Ind. Electron.,
Existing AI methods in power electronic systems are compre- vol. 54, no. 1, pp. 14–33, Feb. 2007.
hensively reviewed in this article. New findings are identified as [15] B. K. Bose, “Artificial intelligence techniques in smart grid and renew-
follows. able energy systems-some example applications,” Proc. IEEE, vol. 105,
no. 11, pp. 2262–2273, Nov. 2017.
1) From the application perspective, the AI methods applied [16] M. Seyedmahmoudian et al., “State of the art artificial intelligence-based
in power electronic systems can be categorized as the MPPT techniques for mitigating partial shading effects on PV systems—
design, control, and maintenance. The usage percentage, A review,” Renewable Sustain. Energy Rev., vol. 64, pp. 435–455,
Oct. 2016.
application trend, features, and requirements of AI in each [17] A. Mellit and S. A. Kalogirou, “Artificial intelligence techniques for
life-cycle phase are discussed. photovoltaic applications: A review,” Progr. Energy Combustion Sci.,
2) From the method perspective, the AI methods applied in vol. 34, no. 5, pp. 574–632, Oct. 2008.
[18] H. S.-H. Chung, H. Wang, F. Blaabjerg, and M. Pecht, Reliability of
power electronic systems can be categorized as expert Power Electronic Converter Systems. London, U.K.: Inst. Eng. Technol.,
system, fuzzy logic, metaheuristic methods, and machine 2015.
learning. The usage percentage, advantages, and limi- [19] M. Riera-Guasp, J. A. Antonino-Daviu, and G. Capolino, “Advances in
electrical machine, power electronic, and drive condition monitoring and
tations of relevant AI algorithms in each category are fault detection: State of the art,” IEEE Trans. Ind. Electron., vol. 62, no. 3,
comprehensively compared. pp. 1746–1759, Mar. 2015.
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4655

[20] H. Soliman, H. Wang, and F. Blaabjerg, “A review of the condition [43] Q. Wang and S. Niu, “Design, modeling, and control of a novel hybrid-
monitoring of capacitors in power electronic converters,” IEEE Trans. excited flux-bidirectional-modulated generator-based wind power gen-
Ind. Appl., vol. 52, no. 6, pp. 4976–4989, Nov./Dec. 2016. eration system,” IEEE Trans. Power Electron., vol. 33, no. 4,
[21] M. Pecht and R. Jaai, “A prognostics and health management roadmap for pp. 3086–3096, Apr. 2018.
information and electronics-rich systems,” Microelectron. Rel., vol. 50, [44] S. Lyden and M. E. Haque, “A simulated annealing global maximum
no. 3, pp. 317–323, Mar. 2010. power point tracking approach for PV modules under partial shading
[22] L. Duchesne, E. Karangelos, and L. Wehenkel, “Recent developments conditions,” IEEE Trans. Power Electron., vol. 31, no. 6, pp. 4171–4181,
in machine learning for energy systems reliability management,” Proc. Jun. 2016.
IEEE, vol. 108, no. 9, pp. 1656–1676, Sep. 2020. [45] J. Zhang, Y. Shi, and Z.-H. Zhan, “Power electronic circuits design: A
[23] R. C. G. J. Pinto and B. Ozpineci, “Tutorial: Artificial intelligence particle swarm optimization approach,” in Proc. Asia-Pac. Conf. Simu-
applications to power electronics,” in Proc. IEEE Energy Convers. Congr. lated Evolution Learn., 2008, pp. 605–614.
Expo., 2019, pp. 1–139. [46] Y. C. Hung, F. J. Lin, J. C. Hwang, J. K. Chang, and K. C. Ruan, “Wavelet
[24] J. Foutz, “Power supply circuit development estimating aid: An expert fuzzy neural network with asymmetric membership function controller
system application,” in Proc. Annu. IEEE Appl. Power Electron. Conf. for electric power steering system via improved differential evolution,”
Expo., 1988, pp. 64–71. IEEE Trans. Power Electron., vol. 30, no. 4, pp. 2350–2362, Apr. 2015.
[25] S. M. Chhaya and B. K. Bose, “Expert system aided automated design, [47] J. Yuan, B. Chen, B. Rao, C. Tian, W. Wang, and X. Xu, “Possible analogy
simulation and controller tuning of AC drive system,” in Proc. 21st Annu. between the optimal digital pulse width modulation technology and the
Conf. IEEE Ind. Electron., vol. 1, 1995, pp. 712–718. equivalent optimisation problem,” IET Power Electron., vol. 5, no. 7,
[26] W. Li and J. P. Ying, “Design and analysis artificial intelligence (AI) pp. 1026–1033, Aug. 2012.
research for power supply-power electronics expert system (PEES),” in [48] S. Singh and B. Singh, “Optimized passive filter design using modi-
Proc. Annu. IEEE Appl. Power Electron. Conf. Expo., vol. 1–4, 2008, fied particle swarm optimization algorithm for a 12-pulse converter-fed
pp. 2009–2015. LCI-synchronous motor drive,” IEEE Trans. Ind. Appl., vol. 50, no. 4,
[27] D. Fezzani, H. Piquet, and H. Foch, “Expert system for the CAD in power pp. 2681–2689, Jul./Aug. 2014.
electronics-application to UPS,” IEEE Trans. Power Electron., vol. 12, [49] B. Zhao, X. Zhang, and J. J. Huang, “AI algorithm-based two-stage opti-
no. 3, pp. 578–586, May 1997. mal design methodology of high-efficiency CLLC resonant converters for
[28] A. M. Elsaadawi, A. E. Kalas, and M. Fawzi, “Development of an expert the hybrid AC-DC microgrid applications,” IEEE Trans. Ind. Electron.,
system to fault diagnosis of three phase induction motor drive system,” vol. 66, no. 12, pp. 9756–9767, Dec. 2019.
in Proc. Int. Middle-East Power Syst. Conf., 2008, pp. 497–502. [50] F. J. Lin, L. T. Teng, and M. H. Yu, “Radial basis function network con-
[29] Y. Izuno, R. Takeda, and M. Nakaoka, “New fuzzy reasoning-based high- trol with improved particle swarm optimization for induction generator
performance speed/position control schemes for ultrasonic motor driven system,” IEEE Trans. Power Electron., vol. 23, no. 4, pp. 2157–2169,
by two-phase resonant inverter,” in Proc. IEEE Ind. Appl. Soc. Annu. Jul. 2008.
Meeting, 1990, pp. 325–330. [51] W. Wang et al., “Training neural-network-based controller on distributed
[30] M. G. Simoes, B. K. Bose, and R. J. Spiegel, “Design and performance machine learning platform for power electronics systems,” in Proc. IEEE
evaluation of a fuzzy-logic-based variable-speed wind generation sys- Energy Convers. Congr. Expo., 2017, pp. 3083–3089.
tem,” IEEE Trans. Ind. Appl., vol. 33, no. 4, pp. 956–965, Jul./Aug. 1997. [52] W. M. Lin and C. M. Hong, “A new Elman neural network-based control
[31] R. Osorio et al., “Fuzzy logic control with an improved algorithm for algorithm for adjustable-pitch variable-speed wind-energy conversion
integrated LED drivers,” IEEE Trans. Ind. Electron., vol. 65, no. 9, systems,” IEEE Trans. Power Electron., vol. 26, no. 2, pp. 473–481,
pp. 6994–7003, Sep. 2018. Feb. 2011.
[32] A. Bubshait and M. G. Simoes, “Design of fuzzy logic-based dynamic [53] F. J. Lin, L. T. Teng, and H. Chu, “A robust recurrent wavelet neural
droop controller of wind turbine system for primary frequency support,” network controller with improved particle swarm optimization for linear
in Proc. IEEE Ind. Appl. Soc. Annu. Meeting, 2018, pp. 1–7. synchronous motor drive,” IEEE Trans. Power Electron., vol. 23, no. 6,
[33] W. Q. Chen and A. M. Bazzi, “Logic-based methods for intelligent pp. 3067–3078, Nov. 2008.
fault diagnosis and recovery in power electronics,” IEEE Trans. Power [54] M. A. Hassan and M. A. Abido, “Optimal design of microgrids in au-
Electron., vol. 32, no. 7, pp. 5573–5589, Jul. 2017. tonomous and grid-connected modes using particle swarm optimization,”
[34] M. G. Simoes and A. Bubshait, “Frequency support of smart grid using IEEE Trans. Power Electron., vol. 26, no. 3, pp. 755–769, Mar. 2011.
fuzzy logic-based controller for wind energy systems,” Energies, vol. 12, [55] K. Ishaque, Z. Salam, M. Amjad, and S. Mekhilef, “An improved particle
no. 8, pp. 1–15, Apr. 2019. swarm optimization (PSO)-based MPPT for PV with reduced steady-state
[35] B. N. Alajmi, K. H. Ahmed, S. J. Finney, and B. W. Williams, “Fuzzy- oscillation,” IEEE Trans. Power Electron., vol. 27, no. 8, pp. 3627–3638,
logic-control approach of a modified hill-climbing method for maximum Aug. 2012.
power point in microgrid standalone photovoltaic system,” IEEE Trans. [56] H. Taghizadeh and M. T. Hagh, “Harmonic elimination of cascade
Power Electron., vol. 26, no. 4, pp. 1022–1030, Apr. 2011. multilevel inverters with nonequal DC sources using particle swarm
[36] C. L. Tseng, S. Y. Wang, S. C. Chien, and C. Y. Chang, “Development of optimization,” IEEE Trans. Ind. Electron., vol. 57, no. 11, pp. 3678–3684,
a self-tuning TSK-Fuzzy speed control strategy for switched reluctance Nov. 2010.
motor,” IEEE Trans. Power Electron., vol. 27, no. 4, pp. 2141–2152, [57] W. Wang, A. C. Liu, H. S. Chung, R. W. Lau, J. Zhang, and A. W.
Apr. 2012. Lo, “Fault diagnosis of photovoltaic panels using dynamic current-
[37] R. C. Garcia, W. I. Suemitsu, and J. O. P. Pinto, “Takagi-Sugeno fuzzy voltage characteristics,” IEEE Trans. Power Electron., vol. 31, no. 2,
model and control of a boost converter using Type-I internal model pp. 1588–1599, Feb. 2016.
control,” in Proc. 39th Annu. Conf. IEEE Ind. Electron. Soc., 2013, [58] Z. H. Liu, H. L. Wei, Q. C. Zhong, K. Liu, X. S. Xiao, and L. H. Wu,
pp. 3794–3799. “Parameter estimation for VSI-fed PMSM based on a dynamic PSO
[38] F. Zidani, D. Diallo, M. E. H. Benbouzid, and R. Nait-Said, “A fuzzy- with learning strategies,” IEEE Trans. Power Electron., vol. 32, no. 4,
based approach for the diagnosis of fault modes in a voltage-fed PWM pp. 3154–3165, Apr. 2017.
inverter induction motor drive,” IEEE Trans. Ind. Electron., vol. 55, no. 2, [59] M. Liserre, A. Dell’Aquila, and F. Blaabjerg, “Genetic algorithm-based
pp. 586–593, Feb. 2008. design of the active damping for an LCL-filter three-phase active recti-
[39] R. Fullér, Introduction to Neuro-Fuzzy Systems. Berlin, Germany: fier,” IEEE Trans. Power Electron., vol. 19, no. 1, pp. 76–86, Jan. 2004.
Springer, 2000, vol. 2. [60] B. Ji, X. G. Song, E. Sciberras, W. P. Cao, Y. H. Hu, and V. Pickert,
[40] Matlab, “Manual of neural network toolbox,” 2005. [Online]. Available: “Multiobjective design optimization of IGBT power modules consider-
http://matlab.izmiran.ru/help/toolbox/nnet/ ing power cycling and thermal cycling,” IEEE Trans. Power Electron.,
[41] J. Zhang, H. S. H. Chung, W. L. Lo, S. Y. Hui, and A. K. M. Wu, vol. 30, no. 5, pp. 2493–2504, May 2015.
“Implementation of a decoupled optimization technique for design of [61] F. Wang, G. Chen, D. Boroyevich, S. Ragon, M. Arpilliere, and V. R.
switching regulators using genetic algorithms,” IEEE Trans. Power Elec- Stefanovic, “Analysis and design optimization of diode front-end rectifier
tron., vol. 16, no. 6, pp. 752–763, Nov. 2001. passive components for voltage source inverters,” IEEE Trans. Power
[42] L. L. Jiang, D. L. Maskell, and J. C. Patra, “A novel ant colony Electron., vol. 23, no. 5, pp. 2278–2289, Sep. 2008.
optimization-based maximum power point tracking for photovoltaic [62] M. D’Antonio, C. Shi, B. Wu, and A. Khaligh, “Design and optimization
systems under partially shaded conditions,” Energy Buildings, vol. 58, of a solar power conversion system for space applications,” IEEE Trans.
pp. 227–236, Mar. 2013. Ind. Appl., vol. 55, no. 3, pp. 2310–2319, May/Jun. 2019.
4656 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

[63] B. Chen, X. Liang, and N. Wan, “Design methodology for inductor- [85] M. Novak and T. Dragicevic, “Supervised imitation learning of finite set
integrated Litz-wired high-power medium-frequency transformer with model predictive control systems for power electronics,” IEEE Trans.
the nanocrystalline core material for isolated DC-link stage of solid- Ind. Electron., to be published, doi: 10.1109/TIE.2020.2969116.
state transformer,” IEEE Trans. Power Electron., vol. 35, no. 11, [86] H. Soliman, P. Davari, H. Wang, and F. Blaabjerg, “Capacitance esti-
pp. 11557–11573, Nov. 2020. mation algorithm based on DC-link voltage harmonics using artificial
[64] L. G. Junior, J. O. P. Pinto, J. A. B. Filho, and G. Lambert-Torres, “Re- neural network in three-phase motor drive systems,” in Proc. IEEE Energy
cursive least square and genetic algorithm based tool for PID controllers Convers. Congr. Expo., 2017, pp. 5795–5802.
tuning,” in Proc. Int. Conf. Intell. Sys. Appl. Power Syst., 2007, pp. 1–6. [87] H. Soliman, H. Wang, B. Gadalla, and F. Blaabjerg, “Condition monitor-
[65] M. J. Schutten and D. A. Torrey, “Genetic algorithms for control of power ing for DC-link capacitors based on artificial neural network algorithm,”
converters,” in Proc. 26th Annu. IEEE Power Electron. Specialists Conf., in Proc. IEEE 5th Int. Conf. Power Eng., Energy Elect. Drives, 2015,
vol. I–II, 1995, pp. 1321–1326. pp. 587–591.
[66] M. S. A. Dahidah and V. G. Agelidis, “Selective harmonic elimina- [88] H. Soliman, I. Abdelsalam, H. Wang, and F. Blaabjerg, “Artificial neural
tion PWM control for cascaded multilevel voltage source converters: network based DC-link capacitance estimation in a diode-bridge front-
A generalized formula,” IEEE Trans. Power Electron., vol. 23, no. 4, end inverter system,” in Proc. IEEE 3rd Int. Future Energy Electron.
pp. 1620–1630, Jul. 2008. Conf. ECCE Asia, 2017, pp. 196–201.
[67] M. J. Neath, A. K. Swain, U. K. Madawala, and D. J. Thrimawithana, “An [89] Z. J. Huang, Z. S. Wang, and H. G. Zhang, “Multilevel feature moving
optimal PID controller for a bidirectional inductive power transfer system average ratio method for fault diagnosis of the microgrid inverter switch,”
using multiobjective genetic algorithm,” IEEE Trans. Power Electron., IEEE-CAA J. Autom. Sin., vol. 4, no. 2, pp. 177–185, Apr. 2017.
vol. 29, no. 3, pp. 1523–1531, Mar. 2014. [90] Z. J. Huang, Z. S. Wang, and H. G. Zhang, “A diagnosis algorithm
[68] A. Mehrizi-Sani and S. Filizadeh, “An optimized space vector modula- for multiple open-circuited faults of microgrid inverters based on main
tion sequence for improved harmonic performance,” IEEE Trans. Ind. fault component analysis,” IEEE Trans. Energy Convers., vol. 33, no. 3,
Electron., vol. 56, no. 8, pp. 2894–2903, Aug. 2009. pp. 925–937, Sep. 2018.
[69] M. Garaj, K. Y. Hong, H. S.-H. Chung, J. Zhou, and A. W.-L. Lo, [91] Z. J. Huang, Z. S. Wang, and H. G. Zhang, “Multiple open-circuit fault
“Photovoltaic panel health diagnostic system for solar power plants,” diagnosis based on multistate data processing and subsection fluctuation
in Proc. IEEE Energy Convers. Congr. Expo., 2018, pp. 1–6. analysis for photovoltaic inverter,” IEEE Trans. Instrum. Meas., vol. 67,
[70] J. Zhang, H. S. H. Chung, A. W. L. Lo, and T. Huang, “Extended ant no. 3, pp. 516–526, Mar. 2018.
colony optimization algorithm for power electronic circuit design,” IEEE [92] S. Qiang and Y. Li, “Motor inverter fault diagnosis using wavelets neural
Trans. Power Electron., vol. 24, no. 1/2, pp. 147–162, Jan. 2009. networks,” in Proc. IEEE Int. Conf. Syst., Man, Cybern., 2013, pp. 3168–
[71] F. J. Lin, P. K. Huang, H. C. Wang, and L. T. Teng, “An induction generator 3173.
system using fuzzy modeling and recurrent fuzzy neural network,” IEEE [93] S. Mohagheghi, R. G. Harley, T. G. Habetler, and D. Divan, “Con-
Trans. Power Electron., vol. 22, no. 1, pp. 260–271, Jan. 2007. dition monitoring of power electronic circuits using artificial neural
[72] B. X. Li and K. S. Low, “Low sampling rate online parameters monitoring networks,” IEEE Trans. Power Electron., vol. 24, no. 10, pp. 2363–2367,
of DC-DC converters for predictive-maintenance using biogeography- Oct. 2009.
based optimization,” IEEE Trans. Power Electron., vol. 31, no. 4, [94] S. Khomfoi and L. M. Tolbert, “Fault diagnostic system for a multilevel
pp. 2870–2879, Apr. 2016. inverter using a neural network,” IEEE Trans. Power Electron., vol. 22,
[73] Q. Sun, Y. Wang, Y. Jiang, and L. Shao, “Non-invasive condition mon- no. 3, pp. 1062–1069, May 2007.
itoring for boost converter based on crow search algorithm,” J. Intell. [95] Y. Cui, J. Shi, and Z. Wang, “Quantum assimilation-based state-of-
Fuzzy Syst., vol. 34, no. 6, pp. 3661–3670, Jun. 2018. health assessment and remaining useful life estimation for electronic
[74] L. Wang, J. Yue, Y. Su, F. Lu, and Q. Sun, “A novel remaining useful life systems,” IEEE Trans. Ind. Electron., vol. 63, no. 4, pp. 2379–2390,
prediction approach for superbuck converter circuits based on modified Apr. 2016.
grey wolf optimizer-support vector regression,” Energies, vol. 10, no. 4, [96] P. Tamilselvan, P. F. Wang, and M. Pecht, “A multi-attribute classifi-
Apr. 2017, Art. no. 459. cation fusion system for insulated gate bipolar transistor diagnostics,”
[75] A. Kavousi, B. Vahidi, R. Salehi, M. K. Bakhshizadeh, N. Farokhnia, Microelectron. Rel., vol. 53, no. 8, pp. 1117–1129, Aug. 2013.
and S. H. Fathi, “Application of the bee algorithm for selective harmonic [97] Z. S. Wang, Z. J. Huang, C. H. Song, and H. G. Zhang, “Multiscale
elimination strategy in multilevel inverters,” IEEE Trans. Power Elec- adaptive fault diagnosis based on signal symmetry reconstitution pre-
tron., vol. 27, no. 4, pp. 1689–1696, Apr. 2012. processing for microgrid inverter under changing load condition,” IEEE
[76] M. H. Etesami, N. Farokhnia, and S. H. Fathi, “Colonial competitive Trans. Smart Grid, vol. 9, no. 2, pp. 797–806, Mar. 2018.
algorithm development toward harmonic minimization in multilevel [98] D. Diallo, M. E. H. Benbouzid, D. Hamad, and X. Pierre, “Fault detection
inverters,” IEEE Trans. Ind. Informat., vol. 11, no. 2, pp. 459–466, and diagnosis in an induction machine drive: A pattern recognition
Apr. 2015. approach based on Concordia stator mean current vector,” IEEE Trans.
[77] K. Haghdar, “Optimal DC source influence on selective harmonic elim- Energy Convers., vol. 20, no. 3, pp. 512–519, Sep. 2005.
ination in multilevel inverters using teaching-learning-based optimiza- [99] C. S. Chen, “TSK-type self-organizing recurrent-neural-fuzzy control of
tion,” IEEE Trans. Ind. Electron., vol. 67, no. 2, pp. 942–949, Feb. 2020. linear microstepping motor drives,” IEEE Trans. Power Electron., vol. 25,
[78] J. Zhang et al., “Evolutionary computation meets machine learning: A no. 9, pp. 2253–2265, Sep. 2010.
survey,” IEEE Comput. Intell. Mag., vol. 6, no. 4, pp. 68–75, Nov. 2011. [100] R. J. Wai and L. C. Shih, “Adaptive fuzzy-neural-network design for
[79] D. Chiozzi, M. Bernardoni, N. Delmonte, and P. Cova, “A neural voltage tracking control of a DC-DC boost converter,” IEEE Trans. Power
network based approach to simulate electrothermal device interaction Electron., vol. 27, no. 4, pp. 2104–2115, Apr. 2012.
in SPICE environment,” IEEE Trans. Power Electron., vol. 34, no. 5, [101] P. Z. Grabowski, M. P. Kazmierkowski, B. K. Bose, and F. Blaabjerg, “A
pp. 4703–4710, May 2019. simple direct-torque neuro-fuzzy control of PWM-inverter-fed induction
[80] T. Dragicevic, P. Wheeler, and F. Blaabjerg, “Artificial intelligence aided motor drive,” IEEE Trans. Ind. Electron., vol. 47, no. 4, pp. 863–870,
automated design for reliability of power electronic systems,” IEEE Aug. 2000.
Trans. Power Electron., vol. 34, no. 8, pp. 7161–7171, Aug. 2019. [102] T. Kamel, Y. Biletskiy, and L. Chang, “Capacitor aging detection for the
[81] M. G. Simoes and B. K. Bose, “Neural-network-based estimation of DC filters in the power electronic converters using ANFIS algorithm,” in
feedback signals for a vector controlled induction-motor drive,” IEEE Proc. IEEE 28th Can. Conf. Elect. Comput. Eng., 2015, pp. 663–668.
Trans. Ind. Appl., vol. 31, no. 3, pp. 620–629, May/Jun. 1995. [103] A. Soualhi et al., “Heath monitoring of capacitors and supercapacitors
[82] J. Zhao and B. K. Bose, “Neural-network-based waveform processing using the neo-fuzzy neural approach,” IEEE Trans. Ind. Informat., vol. 14,
and delayless filtering in power electronics and AC drives,” IEEE Trans. no. 1, pp. 24–34, Jan. 2018.
Ind. Electron., vol. 51, no. 5, pp. 981–991, Oct. 2004. [104] C. Schauder, “Adaptive speed identification for vector control of
[83] S. K. Mondal, J. O. P. Pinto, and B. K. Bose, “A neural-network-based induction-motors without rotational transducers,” IEEE Trans. Ind. Appl.,
space-vector PWM controller for a three-level voltage-fed inverter induc- vol. 28, no. 5, pp. 1054–1061, Sep./Oct. 1992.
tion motor drive,” IEEE Trans. Ind. Appl., vol. 38, no. 3, pp. 660–669, [105] L. E. B. da Silva, B. K. Bose, and J. O. P. Pinto, “Recurrent-neural-
May/Jun. 2002. network-based implementation of a programmable cascaded low-pass
[84] C. M. Lin, K. N. Hung, and C. F. Hsu, “Adaptive neuro-wavelet control filter used in stator flux synthesis of vector-controlled induction mo-
for switching power supplies,” IEEE Trans. Power Electron., vol. 22, tor drive,” IEEE Trans. Ind. Electron., vol. 46, no. 3, pp. 662–665,
no. 1, pp. 87–95, Jan. 2007. Jun. 1999.
ZHAO et al.: OVERVIEW OF AI APPLICATIONS FOR POWER ELECTRONICS 4657

[106] X. G. Fu and S. H. Li, “Control of single-phase grid-connected converters [128] S. Q. Zhou, L. W. Zhou, and P. J. Sun, “Monitoring potential defects in
with LCL filters using recurrent neural network and conventional control an IGBT module based on dynamic changes of the gate current,” IEEE
methods,” IEEE Trans. Power Electron., vol. 31, no. 7, pp. 5354–5364, Trans. Power Electron., vol. 28, no. 3, pp. 1479–1487, Mar. 2013.
Jul. 2016. [129] J. Zhang, H. S. Chung, and W. Lo, “Clustering-based adaptive crossover
[107] J. O. P. Pinto, B. K. Bose, and L. E. B. da Silva, “A stator-flux-oriented and mutation probabilities for genetic algorithms,” IEEE Trans. Evol.
vector-controlled induction motor drive with space-vector PWM and flux- Comput., vol. 11, no. 3, pp. 326–335, 2007.
vector synthesis by neural networks,” IEEE Trans. Ind. Appl., vol. 37, [130] E. G. Strangas, S. Aviyente, and S. S. H. Zaidi, “Time-frequency
no. 5, pp. 1308–1318, Sep./Oct. 2001. analysis for efficient fault diagnosis and failure prognosis for interior
[108] P. Xiao, G. K. Venayagamoorthy, K. A. Corzine, and J. Huang, “Re- permanent-magnet AC motors,” IEEE Trans. Ind. Electron., vol. 55,
current neural networks based impedance measurement technique for no. 12, pp. 4191–4199, Dec. 2008.
power electronic systems,” IEEE Trans. Power Electron., vol. 25, no. 2, [131] S. Zhao, V. Makis, S. Chen, and Y. Li, “Health assessment method for
pp. 382–390, Feb. 2010. electronic components subject to condition monitoring and hard failure,”
[109] Y. Zhang, Z. Wang, H. Wang, and F. Blaabjerg, “Artificial intelligence- IEEE Trans. Instrum. Meas., vol. 68, no. 1, pp. 138–150, Jan. 2019.
aided thermal model considering cross-coupling effects,” IEEE Trans. [132] Y. Z. Lu and A. Christou, “Prognostics of IGBT modules based on the
Power Electron., vol. 35, no. 10, pp. 9998–10002, Oct. 2020. approach of particle filtering,” Microelectron. Rel., vol. 92, pp. 96–105,
[110] M. R. Habibi, H. R. Baghaee, T. Dragicevic, and F. Blaabjerg, “Detection Jan. 2019.
of false data injection cyber-attacks in DC microgrids based on recurrent [133] M. Rigamonti, P. Baraldi, A. Alessi, E. Zio, D. Astigarraga, and
neural networks,” IEEE J. Emerg. Sel. Topics Power Electron., to be A. Galarza, “An ensemble of component-based and population-based
published, doi: 10.1109/JESTPE.2020.2968243. self-organizing maps for the identification of the degradation state of
[111] Z. Li, Z. Zheng, and R. Outbib, “A prognostic methodology for power insulated-gate bipolar transistors,” IEEE Trans. Rel., vol. 67, no. 3,
MOSFETs under thermal stress using echo state network and particle pp. 1304–1313, Sep. 2018.
filter,” Microelectron. Rel., vol. 88–90, pp. 350–354, Sep. 2018. [134] N. Femia, G. Spagnuolo, and V. Tucci, “State-space models and order
[112] Z. J. Huang, Z. S. Wang, X. S. Yao, and H. G. Zhang, “Multi-switches reduction for dc-dc switching converters in discontinuous modes,” IEEE
fault diagnosis based on small low-frequency data for voltage-source Trans. Power Electron., vol. 10, no. 6, pp. 640–650, Nov. 1995.
inverters of PMSM drives,” IEEE Trans. Power Electron., vol. 34, no. 7, [135] J. F. Martins, V. F. Pires, C. Lima, and A. J. Pires, “Fault detection
pp. 6845–6857, Jul. 2019. and diagnosis of grid-connected power inverters using PCA and current
[113] Z. Y. Xue, K. S. Xiahou, M. S. Li, T. Y. Ji, and Q. H. Wu, “Diagnosis mean value,” in Proc. 38th Annu. Conf. IEEE Ind. Electron. Soc., 2012,
of multiple open-circuit switch faults based on long short-term memory pp. 5185–5190.
network for DFIG-based wind turbine systems,” IEEE J. Emerg. Sel. [136] W. Chen, L. Zhang, K. Pattipati, A. M. Bazzi, S. Joshi, and E. M. Dede,
Topics Power Electron., vol. 8, no. 3, pp. 2600–2610, Sep. 2020. “Data-driven approach for fault prognosis of SiC MOSFETs,” IEEE
[114] A. Alghassi, S. Perinpanayagam, and M. Samie, “Stochastic RUL calcu- Trans. Power Electron., vol. 35, no. 4, pp. 4048–4062, Apr. 2020.
lation enhanced with TDNN-based IGBT failure modeling,” IEEE Trans. [137] J. Han, M. Kamber, and J. Pei, Data Mining: Concepts Techniques, 3rd
Rel., vol. 65, no. 2, pp. 558–573, Jun. 2016. ed. San Mateo, CA, USA: Morgan Kaufmann, 2012.
[115] S. Kiranyaz, A. Gastli, L. Ben-Brahim, N. Al-Emadi, and M. Gabbouj, [138] M. Glavic, R. Fonteneau, and D. Ernst, “Reinforcement learning for
“Real-time fault detection and identification for MMC using 1-D con- electric power system decision and control: Past considerations and per-
volutional neural networks,” IEEE Trans. Ind. Electron., vol. 66, no. 11, spectives,” IFAC PapersOnLine, vol. 50, no. 1, pp. 6918–6927, Jul. 2017.
pp. 8760–8771, Nov. 2019. [139] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction.
[116] C. Y. Yin, H. Lu, M. Musallam, C. Bailey, and C. M. Johnson, “A Cambridge, MA, USA: MIT Press, 2018.
prognostic assessment method for power electronics modules,” in Proc. [140] P. Kofinas, S. Doltsinis, A. I. Dounis, and G. A. Vouros, “A reinforcement
2nd Electron. Syst.-Integr. Technol. Conf., 2008, pp. 1353–1358. learning approach for MPPT control method of photovoltaic sources,”
[117] B. Cai, Y. Zhao, H. Liu, and M. Xie, “A data-driven fault diagnosis Renewable Energy, vol. 108, pp. 461–473, Aug. 2017.
methodology in three-phase inverters for PMSM drive systems,” IEEE [141] A. S. Elwer, S. A. Wahsh, M. O. Khalil, and A. M. Nur-Eldeen, “Intel-
Trans. Power Electron., vol. 32, no. 7, pp. 5590–5600, Jul. 2017. ligent fuzzy controller using particle swarm optimization for control of
[118] S. S. Moosavi, A. Kazemi, and H. Akbari, “A comparison of various open- permanent magnet synchronous motor for electric vehicle,” in Proc. 29th
circuit fault detection methods in the IGBT-based DC/AC inverter used Annu. Conf. IEEE Ind. Electron. Soc., vol. 1–3, 2003, pp. 1762–1766.
in electric vehicle,” Eng. Failure Anal., vol. 96, pp. 223–235, Feb. 2019. [142] F. Harashima, Y. Demizu, S. Kondo, and H. Hashimoto, “Application of
[119] S. H. Ali, M. Heydarzadeh, S. Dusmez, X. Li, A. S. Kamath, and B. Akin, neutral networks to power converter control,” in Proc. Conf. Rec. IEEE
“Lifetime estimation of discrete IGBT devices based on Gaussian pro- Ind. Appl. Soc. Annu. Meeting, vol. 1, 1989, pp. 1086–1091.
cess,” IEEE Trans. Ind. Appl., vol. 54, no. 1, pp. 395–403, Jan./Feb. 2018. [143] S. Khomfoi and L. M. Tolbert, “Fault diagnosis and reconfiguration for
[120] J. R. Celaya, A. Saxena, S. Saha, and K. Goebel, “Prognostics of power multilevel inverter drive using AI-based techniques,” IEEE Trans. Ind.
MOSFETs under thermal stress accelerated aging using data-driven Electron., vol. 54, no. 6, pp. 2954–2968, Dec. 2007.
and model-based methodologies,” in Annu. Conf. Prognostics Health [144] H. Zhang, J. Zhao, R. Wang, and T. Ma, “Multi-objective reinforcement
Manage. Soc., 2011, pp. 1–10. learning algorithm and its application in drive system,” in Proc. 34th
[121] A. Lidozzi, L. Solero, F. Crescimbini, and A. Di Napoli, “SVM PMSM Annu. Conf. IEEE Ind. Electron. Soc., vol. 1–5, 2008, pp. 225–230.
drive with low resolution hall-effect sensors,” IEEE Trans. Power Elec- [145] G. Bramerdorfer, J. A. Tapia, J. J. Pyrhonen, and A. Cavagnino, “Modern
tron., vol. 22, no. 1, pp. 282–290, Jan. 2007. electrical machine design optimization: Techniques, trends, and best
[122] N. A. Ahmed and A. K. Al-Othman, “Photovoltaic system with voltage- practices,” IEEE Trans. Ind. Electron., vol. 65, no. 10, pp. 7672–7684,
based maximum power point tracking using support vector machine,” in Oct. 2018.
Proc. 5th IEEE Conf. Ind. Electron. Appl., vol. 4, 2010, pp. 539–544. [146] C. Versele, O. Deblecker, and J. Lobry, “Multiobjective optimal choice
[123] F. Mei, N. Liu, H. Y. Miao, Y. Pan, H. Y. Sha, and J. Y. Zheng, “On-line and design of isolated dc-dc power converters,” in Proc. 14th Eur. Conf.
fault diagnosis model for locomotive traction inverter based on wavelet Power Electron. Appl., 2011, pp. 1–10.
transform and support vector machine,” Microelectron. Rel., vol. 88–90, [147] S. Vighetti, J. Ferrieux, and Y. Lembeye, “Optimization and design of
pp. 1274–1280, Sep. 2018. a cascaded DC/DC converter devoted to grid-connected photovoltaic
[124] C. Delpha, H. Chen, and D. Diallo, “SVM based diagnosis of inverter fed systems,” IEEE Trans. Power Electron., vol. 27, no. 4, pp. 2018–2027,
induction machine drive: A new challenge,” in Proc. 38th Annu. Conf. Apr. 2012.
IEEE Ind. Electron. Soc., 2012, pp. 3931–3936. [148] H. Helali, D. Bergogne, H. Morel, and J. B. H. Slama, “Power converter
[125] X. X. Zheng and P. Peng, “Fault diagnosis of wind power converters based design methodology: Uses of multiple objective techniques for optimiza-
on compressed sensing theory and weight constrained AdaBoost-SVM,” tion of a (42/14V) buck converter,” in Proc. 4th Int. Conf. Integr. Power
J. Power Electron., vol. 19, no. 2, pp. 443–453, Mar. 2019. Syst., 2006, pp. 1–5.
[126] A. G. Abo-Khalil and D. Lee, “DC-link capacitance estimation in [149] H. Zhang, F. Mollet, C. Saudemont, and B. Robyns, “Experimental
AC/DC/AC PWM converters using voltage injection,” IEEE Trans. Ind. validation of energy storage system management strategies for a local DC
Appl., vol. 44, no. 5, pp. 1631–1637, Sep./Oct. 2008. distribution system of more electric aircraft,” IEEE Trans. Ind. Electron.,
[127] T. Z. Wang, H. Xu, J. G. Han, E. Elbouchikhi, and M. E. H. Benbouzid, vol. 57, no. 12, pp. 3905–3916, Dec. 2010.
“Cascaded H-bridge multilevel inverter system fault diagnosis using a [150] Y. F. Yin et al., “Observer-based adaptive sliding mode control of NPC
PCA and multiclass relevance vector machine approach,” IEEE Trans. converters: An RBF neural network approach,” IEEE Trans. Power
Power Electron., vol. 30, no. 12, pp. 7006–7018, Dec. 2015. Electron., vol. 34, no. 4, pp. 3831–3841, Apr. 2019.
4658 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 4, APRIL 2021

[151] F. J. T. Filho, L. M. Tolbert, and B. Ozpineci, “Real time selective Shuai Zhao (Member, IEEE) received the B.E.
harmonic minimization for multilevel inverters using genetic algorithm (hons.), M.E., and Ph.D. degrees in information
and artificial neural network angle generation,” in Proc. 7th Int. Power and communication engineering from Northwest-
Electron. Motion Control Conf., vol. 2, 2012, pp. 895–899. ern Polytechnical University, Xi’an, China, in 2011,
[152] K. H. Tan, “Squirrel-cage induction generator system using wavelet petri 2014, and 2018, respectively.
fuzzy neural network control for wind power applications,” IEEE Trans. He is currently a Postdoctoral Researcher with the
Power Electron., vol. 31, no. 7, pp. 5242–5254, Jul. 2016. Center of Reliable Power Electronics, Department
[153] J. S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference sys- of Energy Technology, Aalborg University, Aalborg,
tem,” IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. 665–685, Denmark. From 2014 to 2016, he was a Visiting
May/Jun. 1993. Ph.D. Student with the Department of Mechanical
[154] S. Zhao, V. Makis, S. Chen, and Y. Li, “Evaluation of reliability function and Industrial Engineering, University of Toronto,
and mean residual life for degrading systems subject to condition moni- Toronto, ON, Canada, with the scholarship from China Scholarship Council. In
toring and random failure,” IEEE Trans. Rel., vol. 67, no. 1, pp. 13–25, August 2018, he was a Visiting Scholar with the Power Electronics and Drives
Mar. 2018. Laboratory, Department of Electrical and Computer Science, The University of
[155] S. Zhao, V. Makis, S. Chen, and Y. Li, “Health evaluation method for Texas at Dallas, Richardson, TX, USA. His research interests include system
degrading systems subject to dependent competing risks,” J. Syst. Eng. informatics, intelligent condition monitoring, diagnostics and prognostics, and
Electron., vol. 29, no. 2, pp. 436–444, Apr. 2018. tailored AI tools for power electronic systems.
[156] IEEE, IEEE Standard Framework for Prognostics and Health Manage-
ment of Electronic Systems, IEEE Standard 1856-2017, pp. 1–31, 2017.
[157] S. Yang, D. Xiang, A. Bryant, P. Mawby, L. Ran, and P. Tavner, “Con-
Frede Blaabjerg (Fellow, IEEE) received the Ph.D.
dition monitoring for device reliability in power electronic converters: A
degree in electrical engineering from Aalborg Uni-
review,” IEEE Trans. Power Electron., vol. 25, no. 11, pp. 2734–2752,
versity, Aalborg, Denmark, in 1995.
Nov. 2010.
He became an Assistant Professor in 1992, an
[158] H. Oh, B. Han, P. McCluskey, C. Han, and B. D. Youn, “Physics-of-
failure, condition monitoring, and prognostics of insulated gate bipolar Associate Professor in 1996, and a Full Professor of
Power Electronics and Drives in 1998. Since 2017,
transistor modules: A review,” IEEE Trans. Power Electron., vol. 30,
he has been a Villum Investigator. He is an Hon-
no. 5, pp. 2413–2426, May 2015.
oris Causa with University Politehnica Timisoara,
[159] M. Al-Greer, M. Armstrong, M. Ahmeid, and D. Giaouris, “Advances
on system identification techniques for DC-DC switch mode power Timisoara, Romania, and Tallinn Technical Univer-
sity, Tallinn, Estonia. He was with ABB-Scandia,
converter applications,” IEEE Trans. Power Electron., vol. 34, no. 7,
Randers, Denmark, from 1987 to 1988. He has au-
pp. 6973–6990, Jul. 2019.
thored/coauthored more than 600 journal papers in the fields of power electronics
[160] S. Zhao, S. Chen, F. Yang, E. Ugur, B. Akin, and H. Wang, “A composite
failure precursor for condition monitoring and remaining useful life and its applications. He is the co-author of four monographs and editor of ten
books in power electronics and its applications. His current research interests
prediction of discrete power devices,” IEEE Trans. Ind. Informat., to
include power electronics and its applications, such as in wind turbines, PV
be published, doi: 10.1109/TII.2020.2991454.
systems, reliability, harmonics, and adjustable speed drives.
[161] MATLAB, “Diagnostic feature designer,” 2020. [Online]. Avail-
able: https://www.mathworks.com/help/predmaint/gs/explore-features- Dr. Blaabjerg is a recipient of 32 IEEE Prize Paper Awards, the IEEE PELS
Distinguished Service Award in 2009, the EPE-PEMC Council Award in 2010,
in-diagnos% tic-feature-designer.html
the IEEE William E. Newell Power Electronics Award 2014, the Villum Kann
[162] M. A. F. Pimentel, D. A. Clifton, L. Clifton, and L. Tarassenko, “A review
Rasmussen Research Award 2014, the Global Energy Prize in 2019, and the 2020
of novelty detection,” Signal Process., vol. 99, pp. 215–249, Jun. 2014.
IEEE Edison Medal. He was the Editor-in-Chief for the IEEE TRANSACTIONS ON
[163] R. Liu, B. Yang, E. Zio, and X. Chen, “Artificial intelligence for fault
diagnosis of rotating machinery: A review,” Mech. Syst. Signal Process., POWER ELECTRONICS from 2006 to 2012. He has been a Distinguished Lecturer
vol. 108, pp. 33–47, Aug. 2018. for the IEEE Power Electronics Society from 2005 to 2007 and for the IEEE
[164] MATLAB, “Predictive maintenance toolbox,” 2020. [Online]. Available: Industry Applications Society from 2010 to 2011 as well as from 2017 to 2018.
https://www.mathworks.com/products/predictive-maintenance.html For the period 2019–2020, he was the President of the IEEE Power Electronics
[165] A. Hanif, Y. Yu, D. DeVoto, and F. Khan, “A comprehensive re- Society. He is also the Vice-President of the Danish Academy of Technical
view toward the state-of-the-art in failure and lifetime predictions of Sciences. He is nominated in 2014–2019 by Thomson Reuters to be between
power electronic devices,” IEEE Trans. Power Electron., vol. 34, no. 5, the 250 most-cited researchers in engineering in the world.
pp. 4729–4746, May 2019.
[166] M. A. Eleffendi and C. M. Johnson, “In-service diagnostics for wire-
bond lift-off and solder fatigue of power semiconductor packages,” IEEE Huai Wang (Senior Member, IEEE) received the B.E.
Trans. Power Electron., vol. 32, no. 9, pp. 7187–7198, Sep. 2017. degree in electrical engineering from the Huazhong
[167] M.-F. Ng, J. Zhao, Q. Yan, G. J. Conduit, and Z. W. Seh, “Predicting University of Science and Technology, Wuhan,
the state of charge and health of batteries using data-driven machine China, in 2007, and the Ph.D. degree in power elec-
learning,” Nature Mach. Intell., vol. 2, no. 3, pp. 161–170, Mar. 2020. tronics from the City University of Hong Kong, Hong
[168] X.-S. Si, W. Wang, C.-H. Hu, and D.-H. Zhou, “Remaining useful life Kong, in 2012.
estimation—A review on the statistical data driven approaches,” Eur. J. He is currently a Professor with the Center of
Oper. Res., vol. 213, no. 1, pp. 1–14, Aug. 2011. Reliable Power Electronics, Department of Energy
[169] Alpha and Omega Semiconductor, “Power semiconductor reliability Technology, Aalborg University, Aalborg, Denmark.
handbook,” Sunnyvale, CA, USA, pp. 8–10, May 2010. [Online]. Avail- He was a Visiting Scientist with the ETH Zurich,
able: http://www.aosmd.com/media/reliability-handbook.pdf Zurich, Switzerland, from August to September 2014,
[170] R. Zhao, R. Q. Yan, Z. H. Chen, K. Z. Mao, P. Wang, and R. X. Gao, and with the Massachusetts Institute of Technology, Cambridge, MA, USA,
“Deep learning and its applications to machine health monitoring,” Mech. from September to November 2013. He was with the ABB Corporate Research
Syst. Signal Process., vol. 115, pp. 213–237, Jan. 2019. Center, Baden, Switzerland, in 2009. He also leads a project on light AI for
[171] Q. Wang, G. Michau, and O. Fink, “Domain adaptive transfer learning cognitive power electronics. His research addresses the fundamental challenges
for fault diagnosis,” in Proc. Prognostics Syst. Health Manage. Conf., in modeling and validation of power electronic component failure mechanisms,
2019, pp. 1–7. and application issues in system-level predictability, condition monitoring,
[172] L. Liao, “Discovering prognostic features using genetic programming in circuit architecture, and robustness design.
remaining useful life prediction,” IEEE Trans. Ind. Electron., vol. 61, Dr. Wang received the Richard M. Bass Outstanding Young Power Electronics
no. 5, pp. 2464–2472, May 2014. Engineer Award from the IEEE Power Electronics Society in 2016, and the
[173] European Union, “General data protection regulation (GDPR),” 2018. Green Talents Award from the German Federal Ministry of Education and
[Online]. Available: https://gdpr.eu/ Research in 2014. He is currently the Chair of IEEE PELS/IAS/IES Chapter
[174] Q. Yang, Y. Liu, T. Chen, and Y. Tong, “Federated machine learning: in Denmark. He is also an Associate Editor for IET Electronics Letters, IEEE
Concept and applications,” ACM Trans. Intell. Syst. Technol., vol. 10, JOURNAL OF EMERGING AND SELECTED TOPICS IN POWER ELECTRONICS, and
no. 2, Jan. 2019, Art. no. 12. IEEE TRANSACTIONS ON POWER ELECTRONICS.