Received 14 May 2023; revised 21 September 2023; accepted 27 November 2023.

Date of publication 1 December 2023;

date of current version 21 December 2023. The review of this article was arranged by Associate Editor J.-L. Schanen.
Digital Object Identifier 10.1109/OJIA.2023.3338534

Artificial Intelligence Techniques for

Enhancing the Performance of Controllers in
Power Converter-Based Systems—An Overview
YUAN GAO 1 (Member, IEEE), SONGDA WANG 2 (Member, IEEE),
TOMISLAV DRAGICEVIC 3 (Senior Member, IEEE), PATRICK WHEELER 4 (Fellow, IEEE),
AND PERICLE ZANCHETTA 4,5 (Fellow, IEEE)
1
School of Engineering, University of Leicester, LE1 7RH Leicester, U.K.
2
Shell Global Solutions International B.V., 2596 HR Den Haag, The Netherlands
3
Department of Energy Technology, Technical University of Denmark, 2800 Kongens, Denmark
4
Power Electronics Machines and Control (PEMC) Group, University of Nottingham, NG7 2RD Nottingham, U.K.
5
Department of Electrical, Computer and Biomedical Engineering, University of Pavia, 27100 Pavia PV, Italy
CORRESPONDING AUTHOR: YUAN GAO (e-mail: [email protected]).
This work was supported by Marie Skłodowska-Curie Actions (MSCA) through the European Union’s Horizon 2020 Research and Innovation Staff Exchange
(RISE) under Grant 872001.

ABSTRACT The integration of artificial intelligence (AI) techniques in power converter-based systems has
the potential to revolutionize the way these systems are optimized and controlled. With the rapid advance-
ments in AI and machine learning technologies, this article presents the analysis and evaluation of these
powerful tools as well as in computational capabilities of microprocessors that control the converter. This
article provides an overview of AI-based controllers, with a focus on online/offline supervised, unsupervised,
and reinforcement-trained controllers. These controllers can be used to create surrogates for inner control
loops, complete power converter controllers, and external supervisory or energy management control. The
benefits of using AI-based controllers are discussed. AI-based controllers reduce the need for complex math-
ematical modeling and enable near-optimal real-time operation via computational efficiency. This can lead to
increased efficiency, reliability, and scalability of power converter-based systems. By using physics-informed
methods, a deeper understanding of the underlying physical processes in power converters can be achieved
and the control performance can be made more robust. Finally, by using data-driven methods, the vast
amounts of data generated by power converter-based systems can be leveraged to analyze the behavior of
the surrounding system and thereby forming the basis for adaptive control. This article discusses several
other potential disruptive impacts that AI could have on a wide variety of power converter-based systems.

INDEX TERMS Artificial intelligence (AI), energy management, machine learning (ML), neural network
(NN), power converter control, renewable energy integration.

I. INTRODUCTION as expert systems, neural networks (NN), and fuzzy logic sys-
The integration of artificial intelligence (AI) techniques in tems. These techniques have been applied to a wide range of
power electronic systems has been an active area of research power electronic systems, including power converters, active
for several decades. The earliest attempts to integrate AI rectifiers, inverters, and motor drives. However, despite the
in power electronic systems focused on rule-based systems, significant progress made in this field, there are still many
where a set of predefined rules were used to control the sys- challenges that need to be addressed to fully realize the po-
tem. However, these rule-based systems were limited in their tential of AI in power electronic systems.
ability to adapt to changing conditions and had difficulty han- Recently, there has been an interest in the use of
dling more complex systems. Over the years, AI techniques AI techniques in power electronic systems, driven by
have evolved to include more sophisticated approaches, such the rapid advancements in machine learning (ML). These
© 2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see
366 https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 4, 2023
 design, control, and maintenance phases were not clearly cate-
gorized and reinforcement learning (RL) was not considered.
For microgrid applications, the survey of AI techniques can
be found in [10] and [11]. Mohammadi et al. [10] discussed
the AI applications in different aspects of microgrids: energy
management, load/generation power forecasting, power con-
verter control, fault detection, cyber attacks, and protection
schemes. Arwa and Folly [11] reviewed the RL-enabled
power management in grid-tide microgrids. Also focusing on
RL, Cao et al. [12] reviewed AI applications in power and
energy systems, based on the introduction of RL principles
and categorization.
Due to the specific control features and challenges of
power converter-based systems, such as high-speed switch-
FIGURE 1. Diagram of AI applications in power electronics. ing, complex modulation, and high-computational burden,
AI implementation in the control of these systems will be
different from the other two life-cycle phases, namely de-
advancements have been made both in the algorithm area sign, and maintenance. Therefore, there is an urgent need
and in the computational capabilities of power converter con- for an analysis of AI applications in the control field for
trollers, which led to a significant improvement in the ability power converter-based systems. In recent years, dramatically
of AI systems to handle power electronic systems. The de- increasing attention has been paid to AI applications in the
sign, control, and preventative maintenance are three key areas control of both individual converters and converter-based
where AI is currently being used in power converter-based (micro)grid systems. This article presents the analysis and
systems [1]. Three parts are illustrated in Fig. 1. summary of the state-of-the-art research on AI and its appli-
The design optimization of converter-based systems usually cation status in enhancing controller performance. Regarding
requires a huge computing power to simulate many feasible the system topology, both inner loop converter control and
design candidates and select the optimal one, especially if (micro)grid-level energy management will be considered in
their fitness needs to be evaluated in various mission profiles. this article.
The main reason for the high computing load is that converter The rest of this article is organized as follows. Section II
needs to be simulated in heat defined design process in itself will list the popular AI techniques (mainly in ML) used in this
is complex in nature where mutually coupled performance target domain. In Sections III and IV, this article will focus
indices, such as efficiency, power density, cost, and reliability, on two aspects of AI-aided controllers: linear and nonlinear.
should be considered [2]. In addition, intelligent control is Both design and control applications are discussed in these
essential for the reliable, robust, and stable operation of power two sections. Furthermore, the specific AI application areas
converter-based systems. For example, when AI algorithms will be reviewed in Section V. Finally, Section VI concludes
are added to the traditional PI/PR or advanced nonlinear this article.
controllers, the dynamic response and tracking performance
may be improved in several aspects, such as reliability to II. AI TECHNIQUES
changing system conditions, reduction of computational load, AI can be implemented using predefined rules, but in most
or consequent performance improvement [3], [4], [5]. Fi- cases, ML algorithms are used to perform specific AI tasks
nally, preventive maintenance of systems, including condition based on learning from the collected data. ML algorithms can
monitoring, fault tolerance, and fault diagnosis, are effective effectively learn rules and relations from training data and
approaches to ensure a system’s healthy operation. With the improve the trained models automatically through experience.
help of AI, the desired prediction and monitoring can be Therefore, the largest use of AI in power converter-based
achieved much faster and more precisely than with traditional systems is with ML. Other common AI methods include
methods [6], [7], [8]. expert systems, fuzzy logic, and metaheuristic methods. To
In [1], a general review of AI algorithms and applications distinguish ML from the metaheuristic search algorithms, Gao
for power electronics was presented from a life-cycle perspec- et al. [13] proposed a simple algorithm categorization that
tive. This article generally considered most AI techniques: comprises search algorithm and surrogate algorithm. Both
expert systems, fuzzy logic, heuristics, and ML, in three dis- categories were used for the same optimization problem for
tinctive life-cycle phases: design, control, and maintenance. a converter-based actuation system.
Therefore, it is not specific to the controller domain in power Three main groups of ML algorithms are supervised learn-
electronics. Supervised and unsupervised learning (UL) tech- ing (SL), UL, and RL [1], [14], as shown in Fig. 2. Generally,
niques for electrical power systems were reviewed in [9]. both SL and UL require collecting data before training. An SL
Although popular ML algorithms were introduced, such as dataset should give outputs/labels corresponding to the inputs,
recurrent NN and random forest (RF), the application areas in while there is no label defined in UL. For RL, there is no

VOLUME 4, 2023 367

GAO ET AL.: AI TECHNIQUES FOR ENHANCING THE PERFORMANCE OF CONTROLLERS IN POWER CONVERTER-BASED SYSTEMS

can also be used to reduce data dimensionality without losing

important information. Therefore, the training data for UL can
only have certain elements. For example, in a flower category
case, there are four elements: sepal length, sepal width, petal
length, and petal width. The UL can discover the information
that was previously undetected. An example is to categorize
different kinds of flowers by using some sampled data of four
elements.
UL problems can be further grouped into clustering and
association problems. Clustering mainly deals with finding
a structure or pattern in unlabeled data. The flower category
case is indeed a clustering problem. Association rules can
discover associations/relationships between elements in large
FIGURE 2. Diagram of ML methods, modes, and online/offline databases. For example, people that buy a new house are most
applications.
likely to buy new furniture.

C. REINFORCEMENT LEARNING
data collection or offline training, because it is determined The third group of ML is RL, which is learning how intelligent
by the trial-and-error exploration of agents in an unknown agents can perform a task by interacting with their environ-
environment in order to maximize the cumulative rewards. ment. Unlike classical and heuristic optimization methods, RL
As shown in Fig. 2, two learning modes in ML are does not require an accurate model of the system or environ-
regression and classification. And in practice, different on- ment to generate an optimal solution [11], [18]. In the learning
line/offline approaches should be investigated for different process, these agents ought to take action (updating the policy
applications. In this article, both online and offline ML model at the same time) in a specific environment. They would get
training are reviewed for the parameter design of controllers; rewards from each step/trying and the final target is usually to
but, for controller imitation learning, a common way is us- maximize the cumulative reward.
ing the collected data for offline training. Based on the Therefore, in the beginning, how to operate in the environ-
input/output model design and the data-driven process, ML ment is unknown, and no trustworthy data can be collected.
can be an effective way to enhance the control performance of The agents should try different actions in the environment to
converter-based systems. Three ML groups will be introduced receive the corresponding rewards. Based on the cumulative
below, then ML-based emerging techniques will be discussed. rewards, the agents can learn a good policy (which means how
to act in a certain position) in the environment. Obviously, RL
A. SUPERVISED LEARNING is an online learning process. It differs from SL or UL because
If there are outputs (also called targets and answers) defined no input/output data pairs are collected before the learning
in the training data of ML, SL methods are usually applied. process.
These targets can either be continuous numbers or integers.
If the targets are all integers/categories, the learning task is D. ML-BASED EMERGING TECHNIQUES
known as classification, otherwise, it would be a regression The above three sections individually introduced three groups
task. Therefore, both regression and classification are training in ML. However, the practical problems may need multiple
the sampled input–output pairs but their output data features groups of algorithms to address, and even an independent
are different. Regression has no requirement for sampled data algorithm could encompass more than one type of learning
but, in classification learning, the training outputs should in- method. For example, deep deterministic policy gradient
clude predefined categories/classes or even only 0 or 1 for usually combined SL [deep neural network (DNN)] with
some specific problems. RL [19], [20].
According to the statistics in [1], usage of SL is 91% of all With the fast development of AI and ML, there are emerg-
ML applications in power electronics. In particular, artificial ing ML techniques that came to worldwide attention in
neural network (ANN) is one of the most common algorithms recent years. Federated learning (FL), also known as col-
in SL where neurons are the fundamental information process- laborative learning, is one of the emerging techniques. It is
ing units and the building blocks. Other SL algorithms include first introduced by Google in 2016 [21]. FL typically ap-
linear regression [15], support vector machine [13], [16], and plies when individual actors need to train models on larger
RF [17]. datasets than their own, but cannot afford to share the data
in itself with others (e.g., for legal, strategic, or economic
B. UNSUPERVISED LEARNING reasons). The technology yet requires good connections be-
For UL, there is no predefined output in the training data. tween local servers and minimum computational power for
UL learns how to discover patterns and information from a each node [22], [23]. FL trains an algorithm across multi-
dataset without preset outputs/labels/features. Such methods ple decentralized edge devices or servers holding local data

368 VOLUME 4, 2023

samples, without exchanging them. This approach stands in
contrast to traditional centralized ML techniques where all the
local datasets are uploaded to one server, as well as to more
classical decentralized approaches, which often assume that
local data samples are identically distributed. There are three
parts to FL: centralized, decentralized, and heterogeneous.
FL has been effectively applied to self-driving [24], wire-
less power control [25], and large-scale energy systems [23].
With the continuing development of distributed large-scale
converted-based systems, FL could become a promising AI
technique for wireless computing and power control.
Another popular technique in ML is transfer learning (TL),
it has found several applications in electrical power systems.
TL leverages knowledge gained from one task and applies
it to a related but different task to pursue high performance
and efficiency. TL is desirable for the practical problems
that the training data are expensive or impossible to recol-
lect [26]. Some application areas of TL in electrical domain
are energy management [27], battery health [28], impedance FIGURE 3. ANN applications in linear controllers.
estimation [29], [30], and thermal performance of motor [31].

In [38], two types of ANNs were proposed to design a

III. AI APPLICATIONS IN LINEAR CONTROLLERS OF Hammerstein model identifier and a PID controller for the
POWER CONVERTERS adaptive control of a converter-based hybrid energy storage
Linear controllers, such as cascaded voltage and current loops, system (HESS). The online learning of Hammerstein ANN
are widely used in power converter systems to regulate the can supply an appropriate power flow reference for the HESS
power flow and maintain stability. However, these controllers to improve unsatisfactory frequency deviations and tie power
are designed to operate in linear systems and can have diffi- oscillation. In [39], the student psychology optimization al-
culty performing optimally in nonlinear systems. Therefore, gorithm is used for the online tuning of control parameters
for optimal performance, the parameters of linear control within a PI-incorporated RL ANN controller, for power sys-
loops should be adapted to the changing operating conditions tem applications.
in nonlinear surrounding environments. In this context, AI- Apart from ANN applications, in [40], local model net-
based techniques can be used to adapt the parameters of linear works (also recognized as neuro-fuzzy systems) were used
controllers in power converters. to identify the model of the dc–dc buck converter, which
Due to the possession of nonlinearity, and the ability of on- can model the nonlinear dynamics for the implementation
line learning (transferring the real-time experience to AI mod- of a local linear controller regulating the converter output
els), the AI-aided PI/PR online control methods have been voltage. To improve the transient response performance of
widely used in different areas, such as dc–dc converter [3], dc–dc converters, Liu et al. [41] combined two fuzzy-logic
[32], [33], [34], spacecraft [35], exoskeleton systems [36], controllers with a PID compensator for the coefficients and
and electric vehicle charging system [37]. However, there are error modifications.
generally two groups of approaches utilized to do that, design
and control, as discussed below.
2) OFFLINE DESIGN
Here, a surrogate model of the optimal parameter adapta-
A. PARAMETER ADAPTATION tion method is created in an offline setting, using data from
1) ONLINE DESIGN the model of the environment and the converter system.
In this approach, an AI-based signal processing method is This model can be used to determine the optimal param-
used to take measurements from the environment and learn eters for different operating conditions, which can then be
the optimal parameters in real time. This can be achieved by implemented in the controller. This approach does not re-
exploring an online AI-based surrogate model of the linear quire real-time adaptation, but it does require a large amount
system, which is trained using data from the system and the of simulation data to be collected and processed in order
environment. AI techniques, such as ANNs and heuristic al- to determine the optimal parameters. Moreover, if simu-
gorithms, can be used to build this model. By continuously lation data are not sufficiently well representing the real
searching through the model, the controller can adapt the system, this approach may yield suboptimal or even unstable
parameters for different operating conditions in real time. results.
However, it can be computationally intensive and require a The online/offline ANN-based design diagrams in a linear
high-frequency sampling rate. controller are shown in Fig. 3. This figure uses a dc–ac voltage

VOLUME 4, 2023 369

GAO ET AL.: AI TECHNIQUES FOR ENHANCING THE PERFORMANCE OF CONTROLLERS IN POWER CONVERTER-BASED SYSTEMS

source inverter (VSI) feeding an ac load via an output LC filter A. PARAMETER DESIGN FOR NONLINEAR CONTROLLERS
but, the general application of ANN should be similar in other Weighting factor design continues to be a hot topic in MPC
topologies, such as ac (micro)grid connected using an LCL because there are usually multiple control terms in the used
filter. The biggest difference between online and offline design CF. Dragičević and Novak [45] proposed an ANN approach
is that there is no intensive online training in offline design. to automatically select the weighting factors in the CF of
As a tradeoff, independent offline learning usually requires FCS-MPC. The trained ANN serves as a surrogate model of
a high number of offline simulations to collect the training the converter that can provide fast and accurate estimates of
data. For dc–dc converters, Liu et al. [32] proposed using the performance metrics for any weighting factor combina-
ANN to adjust both PID coefficients and the controller struc- tion. Vazquez et al. [46] presented an ANN-based real-time
ture. Maruta et al. [3] proposed an ANN-based predictor to tuning method of a weighting factor to achieve the desired
modify the output voltage reference values in a PID controller, average switching frequency and track the current reference.
which can improve the transient response. In these studies, Gao et al. [47] proposed an inverse application method of
ANN was trained offline to pursue high prediction accuracy AI that can effectively provide references and coefficients for
for the output voltage of the dc–dc converter. the control of a power converter-based system. Two different
cases were used for the method validation. One is the current
sharing for a converter-based microgrid. The other is the ex-
B. CONTROL SURROGATE MODELING tension of the MMC operation region under unbalanced grid
Different from the above parameter adaptation studies, in [33], faults. An ANN-based droop coefficient design method was
an ANN was trained as an adaptive controller, which directly proposed in [48] for improved load sharing.
outputs the duty cycle and the frequency of gate-driving sig-
nals. This belongs to the other group of AI applications in the
linear controllers, the controller surrogate model. In this ap- B. IMITATION CONTROLLER FOR NONLINEAR
proach, a surrogate AI-based model of the system is designed CONTROLLERS
that blends linear control and nonlinear system dynamics, e.g., The high-computational burden is one of the main disadvan-
by matching the linear impedance model of the converter with tages of MPC, especially for the implementation of multistep
the nonlinear impedance model of the grid. This model can predictions or/and multilevel converters. To address that,
then be used in an online or offline search for parameters ANNs were used to learn the MPC model via offline training,
of the linear controller, as indicated in previous subsections. which can keep an approximate control performance while at
The AI-based model can be created with techniques, such the same time reducing the computational burden. The theo-
as physics-informed NN, which can incorporate the physical retical basis of this imitation approach is that the predictive
laws of the system into the model. control process is completely deterministic, i.e., for the same
set of input variables (i.e., circuit measurements) and a given
CF, the outputs (inserted vectors/submodules) will always be
IV. AI APPLICATIONS IN NONLINEAR CONTROLLERS OF the same. In this context, while the conventional MPC uses ex-
POWER CONVERTERS haustive rolling optimization every time instant to identify the
If only simple linear controllers are employed, system dy- optimal actuation, this is not necessary. It should be possible
namics and external disturbances may not be well addressed to represent the deterministic input–output relationship with a
though they are computationally light without much complex- more computationally efficient structure. Therefore, the same
ity. In contrast, some advanced control methods can integrate control effect as the MPC can be achieved by an ANN, but
nonlinearities and consider system constraints into the model, with a lower online-computational burden.
for example, model predictive control (MPC). One of the The general way of MPC imitation is depicted in Fig. 4 for
key characteristics of MPC is that the control objectives and an inverter system. There are three steps in this process.
system limitations can be simply and intuitively included in 1) Data collection from the original MPC model.
the cost function (CF), directing the easy generalization for 2) ANN offline training.
different converter topologies. 3) Online test of the trained ANN.
Two main categories in MPC include the continuous control The first two steps are shown in Fig. 4(a) while the last step
set MPC (CCS-MPC) and the finite control set MPC (FCS- is shown in Fig. 4(b).
MPC). CCS-MPC uses the control vector as a continuous Some published works studying ANN-based predictive
control signal, thus, the output of optimization can be any controllers are discussed below. A deep NN-based predictive
vector within the control region defined by available voltage control strategy is generally presented in [49] for the applica-
vectors of the converter [42], [43]. Differently, FCS-MPC tion of high-frequency multilevel converters. However, none
considers a set with a limited number of input candidates thus of the technical details of ML work was provided, for exam-
the output of optimization can only be one of the considered ple, data collection, and ANN training and validation. In con-
vectors in the set [44]. AI techniques have been used for both trast, Mohamed et al. [50] clearly presented an ANN learning
the design and imitation control of MPC. The following two approach for the control of a two-level VSI feeding linear and
sections will look at these two aspects in detail. nonlinear loads. System descriptions, FCS-MPC principles,

370 VOLUME 4, 2023

 algorithm (GA) [57]. For example, Hou et al. [58] proposed
an adaptive self-organizing fuzzy wavelet NN for an ac mo-
tor servo system, which uses adaptive GA to optimize ANN
parameters. In [5] and [59], recurrent fuzzy ANN-aided (for
feature selection) sliding-mode control was used in the power
converter-based systems for intelligent control with an active
power filter, where the online learning laws can effectively
capture unknown functions with a low-computational burden.

V. APPLICATION AREAS AND FUTURE PERSPECTIVES

After discussing the AI techniques for linear and nonlin-
ear controllers, in this section, different AI applications are
brieftly introduced. Fig. 5 shows the general structure of
this section. Solar, wind, high-voltage direct current (HVDC),
energy storage, powertrain, and promising power to X appli-
cations are presented to describe the differences between each
other.
Following the above sections, the common technolo-
gies will be first introduced in this section, those methods
are general methods and thus available for most applica-
tions/situations. Furthermore, application-specific AI tech-
nologies are given one by one, and those technologies are
designed for specific applications to achieve different goals.

A. GENERAL AI TECHNOLOGIES
Four aspects will be summarized below for the general AI
technologies. The first two (AI-based controller and controller
FIGURE 4. Diagram of MPC imitation learning. (a) Data collection and ANN design using AI) have been elaborated in Section IV. The third
training offline. (b) Trained ANN online test. (converter modeling using AI) and fourth (stability analysis
by AI) are also important and practical topics regarding AI-
proposed ANN architecture, and training/test/validation are enabled controller. The pros and cons of each aspect will be
all presented in detail. Moreover, the authors collected sample presented to help the potential scholars and engineers select
data from 60 circuit conditions (covering the linear/nonlinear suitable methods.
load) and tested the trained ANN for all the considered circuit
conditions. Even though the proposed method in [50] was not 1) AI-BASED CONTROLLER
validated in the experiment, it is presented with all necessary The grid-connected converters play important roles to regu-
technical details, thus easy to implement and generalize. Fol- late the power to the grid. Section IV introduced the ANN
lowing the same approach, Novak and Dragicevic [51] trained application in replacing the traditional PI-based controllers to
different ANN controllers for upto N = 3 prediction horizons reduce the computation burden. For the photovoltaics (PV)
and tested them in the experiment rig. application, Demirtas et al. [60] gave an example that used an
Regarding the converter topology, the authors in [50] and ANN-based controller to control a single-phase PV inverter.
[51] used this SL method for two-level converters, and Novak The results show that the output power can be regulated by
and Blaabjerg [52] further implemented it in the three-level the proposed ANN-based controller. Similar AI-based imita-
converter. The authors in [49] and [53] trained and validated tion methods could be generalized to different systems and
the ANN imitation controller to a five-level flying capacitor applications.
converter. The authors in [54] and [55] proposed two dif- The advantage of the AI-based imitation controller is that,
ferent ANN imitation controllers for the modular multilevel the transient response of the AI controller is faster than PI
converter (MMC). controller because there is no integral part in the controller
to slow down the dynamics behavior. However, because the
C. ONLINE ANN FOR NONLINEAR CONTROLLERS controller is trained by collected data only, there may be a risk
Apart from the above design and control studies using ANN of running out of the training range. Thus, the response of AI
offline training for MPC, there was plenty of research on controller is not fully predictable. In [55], a case is given with
NN-based nonlinear online-learning controllers, which do not some results when the controller runs out of the training range,
comprise data collection for the learning but use an optimiza- the MMC is still able to control the system. But, obviously, it
tion method, such as Levenberg–Marquard [56] and genetic is not for all the related control problems.

VOLUME 4, 2023 371

GAO ET AL.: AI TECHNIQUES FOR ENHANCING THE PERFORMANCE OF CONTROLLERS IN POWER CONVERTER-BASED SYSTEMS

FIGURE 5. AI applications.

2) CONTROLLER DESIGN USING AI verification, it may not be applicable in the early design phase
As mentioned in Section IV, apart from replacing the PI/PR of the system.
controllers with AI, AI technologies can be used to help better
design the traditional converters, at the same time, keeping the 4) AI-BASED STABILITY ANALYSIS
existing controllers in the converters. For example, in [61], In the modern power electronics-based power system, the
the artificial bee colony algorithm is used to design the PID stability of the system is an important factor due to the high
parameters of the PV converters. Galotto et al. [62] presented share of inverter-based resources. For example, the phase lock
an recursive least square and GA-based tool for PID controller loop of the grid following the converter will cause system
tuning, with this method, the controller design process can be instability under weak grid conditions. AI technologies can
accelerated. In [63], a particle swarm optimization controller contribute to stabilizing power electronics-based power sys-
parameter design method is proposed for the dc–dc buck con- tems. In [65], a RF-based power oscillation damper for grid
verter to reduce the aforementioned transient and steady-state forming converter is proposed, the proposed damper can au-
errors. tomatically adjust the gain of the active damping controller
For ML applications, Dragičević and Novak [45] proposed based on the different operating points. In [66], the proposed
an ANN-based method to design the weighting factors of the AI-based control design uses a DNN to learn the nonlinear
MPC-based grid-connected controller. In [47], a special appli- mapping between the virtual synchronous generator (VSG)
cation of ANN is introduced to select the droop coefficients input and output signals, enabling it to adapt to different
of microgrids, that is, instead of inputting system parameters operating conditions and disturbances. This article presents
for system response, the desired system response is set as simulation results to demonstrate the effectiveness of the pro-
the input of ANN, then the trained ANN can tell the user posed control design in improving the stability and robustness
what is the preferable input parameters for the target system. of VSGs under various operating scenarios. Liu et al. [67]
The gain of ML approach is that, it can get rid of optimiza- gave a comprehensive review of the AI applications to im-
tion algorithms, no local optimal problem. This is because, prove the stability of future power systems.
after training, computation of ANN is extremely light, thus, For stability analysis, the key interest from the industry is
exhaustive algorithm can be used in the design space. The the physical principles and insights of stability issue in the
optimal design can be quickly found through tightly sweeping system. With traditional transfer function or state-space-based
the design space. stability analysis, there is a solid theory behind it to explain
the stability. However, due to lack of interpretation, industries
are still hesitant to embrace the AI-based stability analysis,
3) CONVERTER MODELING USING AI though it is easy to use with promising results.
Regarding the modeling of the power electronics convert-
ers, traditional analytical methods can be used to model the B. TECHNOLOGIES FOR SPECIFIC APPLICATION
converter, however, it heavily relies on the accuracy of the AI techniques can be applied in many specific areas, but
system parameters. By using AI technologies, the modeling the function and purpose may vary according to different
of the converter can be obtained by measuring the converters. requirements. For example, maximum power point tracking
In [64], Bayesian regularization along with ANN and RF- (MPPT) should be considered for solar applications; however,
based ML methods are used to model the power electronics AI-based MPPT is not very common in powertrain applica-
converter with collected experimental data. tions. Some specific AI application areas are listed as follows.
In this case, most of the industrial applications will still
treat AI converter model as a trained or simulated “black 1) SOLAR
box.” That means, this method can only be used to verify For solar applications, the special application of the AI tech-
the system modeling after the physical system has been built nology is for the MPPT purpose, MPPT is widely used in
up. Therefore, even though this method is useful for modeling dc–dc converters in solar applications to track the maximum

372 VOLUME 4, 2023

power from the sun. Traditionally, two main non-AI MPPT the battery-connected converter can also provide the grid an-
algorithms are commonly used: perturb and observe algo- cillary services, for instance, reducing the system harmonics,
rithm [68] and the incremental conductance algorithm [69]. regulating system frequency, and supporting system voltage.
However, the traditional MPPT algorithms cannot properly Koganti et al. [75] proposed the design of a multi-objective-
track the maximum power point under partial shading con- based AI controller for a wind/battery-connected shunt active
ditions. The AI-added MPPT methods can track the global power filter (SAPF). The controller aims to improve the per-
maximum power point with higher computational speed and formance of the SAPF by simultaneously optimizing multiple
faster dynamic speed [70]. Kiran et al. [71] compared different objectives, such as total harmonic distortion, voltage reg-
AI-aided MPPT algorithms both under partial shading condi- ulation, and energy efficiency. Abdalla et al. [76] gave a
tions, the conclusion from this article is that RBFC optimized comprehensive review of the AI applications of energy storage
Fuzzy controller has the highest efficiency under partial shad- systems.
ing conditions.

2) WIND 5) POWERTRAIN FOR TRANSPORTATION: ELECTRIC DRIVE

CONTROL
In wind energy systems, one challenge is the changing of the
wind speed over time might cause power fluctuation in the Several studies have been conducted for electric drive applica-
wind energy system, then the frequency of the wind network tions by using ML-aided controllers. For example, Hammoud
will also be fluctuating. In [72], an adaptive ANN controller et al. [77] used a multilayer perception feedforward ANN to
for the energy capacitor system in a wind farm is proposed to learn the long-horizon FCS-MPC, which solves the mixed-
better control the frequency of the wind farm network. The integer optimal control problem offline to generate the training
network frequency variation is minimized by the proposed data of ANN. After training, a two-level inverter-based motor
controller. In [73], an insulated-gate bipolar transistor (IGBT) drive was controlled by the trained ANNs in real-time sys-
fault tolerant control strategy for wind turbine converter is tems, and different horizon numbers were compared. Deep
proposed, and fuzzy logic is used to fast detect the wind RL was used in [78] for a dc–dc buck converter-fed dc
turbine converter IGBT fault, then the protection algorithm motor to reduce the torque/current ripples. And Nikdel and
will activate the redundant converter leg to bypass the faulty Nikdel [79] proposed ANN model predictive and variable
leg. structure controllers for the single-degree-of-freedom rotary
manipulator.
3) HVDC APPLICATION
The power needs to be transferred to the power grid using 6) FUTURE PERSPECTIVES—POWER TO HYDROGEN
high-voltage alternate current (HVAC)/HVDC technologies With the large-scale development of renewable energies, the
to reduce transmission losses from the far away generation constraint of the power grid limits the maximum power that
units. For HVAC technology, traditional power transformers can be transferred to the grid, and also the intermittent nature
are used, which will not require many power electronics con- of renewable energies requires a long-term energy storage
trollers in the plant, then AI application in the HVDC system solution for renewable energies. Producing hydrogen by using
is not within the scope of this article. For HVDC technol- energy can be an ideal solution for long-term energy storage.
ogy, two main power electronics converters are used: line- AI has good potential to better control and design future green
commutated HVDC converters, and voltage-source HVDC hydrogen systems. The power converters in green hydrogen
converters. AI technologies can be used to better control the systems play important roles in safe, high-efficient, and low-
HVDC converters to achieve internal and external perfor- cost green hydrogen production. This area is still relatively
mances. In [74], an ANN-based operational region extension new and many questions need to be solved in the near future.
method for MMCs under unbalanced grid faults is proposed, Nowadays, the power industry is still mainly using con-
the ANN can automatically decide the injected circulating ventional analytical methods to design, model, and control
current to reduce the submodule voltage ripple under grid the power electronics system, because of the well-developed
faults, hence, the submodule over voltage trip is avoided and theory and interpretation of the conventional methods. An
then the operating region is extended. In [54] and [55], ML- example is the impedance-based method to judge the grid-
based controllers are proposed for MMCs to achieve faster connected converter stability. In the near future, the power
dynamic response and lower the computation burden of the industry will not fully trust AI techniques, nor quickly turn
controller. to AI-based methods for power converter control. However,
AI-based methods still have great potential for industrial ap-
4) ENERGY STORAGE SYSTEM plications, especially when the system is too complicated to
Energy storage is an important role in the future renewable- be analytically explained. The data-driven models could help
energy-based power system to shift the mismatch between electrical engineers to gain the system insights without doing
generation and consumption. The battery is one of the most tedious analysis or modeling, this is indeed the key advantage
important ways to do energy storage. Besides energy storage, of AI techniques.

VOLUME 4, 2023 373

GAO ET AL.: AI TECHNIQUES FOR ENHANCING THE PERFORMANCE OF CONTROLLERS IN POWER CONVERTER-BASED SYSTEMS

VI. CONCLUSION [14] T. Guillod, P. Papamanolis, and J. W. Kolar, “Artificial neural network
This article gave an overview of AI-enabled controllers for (ANN) based fast and accurate inductor modeling and design,” IEEE
Open J. Power Electron., vol. 1, pp. 284–299, 2020.
power converter-based systems. To begin with, the fast- [15] S. Weisberg, Applied Linear Regression, vol. 528. Hoboken, NJ, USA:
developing ML algorithms were introduced with three basic John Wiley & Sons, Inc., 2005.
groups: supervised, unsupervised, and RL. Then, the AI ap- [16] J. A. Suykens and J. Vandewalle, “Least squares support vector machine
classifiers,” Neural Process. Lett., vol. 9, no. 3, pp. 293–300, 1999.
plications in linear and nonlinear controllers were discussed [17] M. Hannan, J. Abd Ali, A. Mohamed, and M. N. Uddin, “A random
separately from two perspectives: design and control. In this forest regression based space vector PWM inverter controller for the
part, online/offline application modes and their technical de- induction motor drive,” IEEE Trans. Ind. Electron., vol. 64, no. 4,
pp. 2689–2699, Apr. 2017.
tails were analyzed and discussed. In addition, the common [18] L. P. Kaelbling, M. L. Littman, and A. W. Moore, “Reinforcement
AI technologies were summarized with four aspects that can learning: A survey,” J. Artif. Intell. Res., vol. 4, pp. 237–285, 1996.
be generally used for various applications. Some specific AI [19] M. Gheisarnejad, H. Farsizadeh, and M. H. Khooban, “A novel non-
linear deep reinforcement learning controller for DC–DC power buck
application areas were further given, which may focus on converters,” IEEE Trans. Ind. Electron., vol. 68, no. 8, pp. 6849–6858,
typical requirements or targets. Finally, a future perspective on Aug. 2021.
power-to-hydrogen was added as a promising way for future [20] A. J. Abianeh, Y. Wan, F. Ferdowsi, N. Mijatovic, and T. Dragičević,
“Vulnerability identification and remediation of FDI attacks in islanded
green energy applications. DC microgrids using multiagent reinforcement learning,” IEEE Trans.
Power Electron., vol. 37, no. 6, pp. 6359–6370, Jun. 2022.
[21] J. Konečný, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh, and D.
Bacon, “Federated learning: Strategies for improving communication
REFERENCES efficiency,” 2016, arXiv:1610.05492.
[1] S. Zhao, F. Blaabjerg, and H. Wang, “An overview of artificial intelli- [22] P. Kairouz et al., “Advances and open problems in federated learning,”
gence applications for power electronics,” IEEE Trans. Power Electron., Found. Trends Mach. Learn., vol. 14, no. 1–2, pp. 1–210, 2021.
vol. 36, no. 4, pp. 4633–4658, Apr. 2021. [23] X. Cheng, C. Li, and X. Liu, “A review of federated learning in energy
[2] R. Rajamony, S. Wang, G. Calderon-Lopez, I. Ludtke, and W. Ming, systems,” in Proc. IEEE/IAS Ind. Commercial Power Syst. Asia, 2022,
“Artificial neural networks-based multi-objective design methodology pp. 2089–2095.
for wide-bandgap power electronics converters,” IEEE Open J. Power [24] A. M. Elbir, B. Soner, S. Çöleri, D. Gündüz, and M. Bennis, “Federated
Electron., vol. 3, pp. 599–610, 2022. learning in vehicular networks,” in Proc. IEEE Int. Mediterranean Conf.
[3] H. Maruta, H. Taniguchi, and F. Kurokawa, “A study on effects of Commun. Netw., 2022, pp. 72–77.
different control period of neural network based reference modified pid [25] D. Liu and O. Simeone, “Privacy for free: Wireless federated learning
control for DC-DC converters,” in Proc. 15th IEEE Int. Conf. Mach. via uncoded transmission with adaptive power control,” IEEE J. Sel.
Learn. Appl., 2016, pp. 460–465. Areas Commun., vol. 39, no. 1, pp. 170–185, Jan. 2021.
[4] P. Pandey, S. Bodda, and P. Agnihotri, “Comparative analysis of the [26] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans.
control of LLC resonant converter using proportional-integral and ma- Knowl. Data Eng., vol. 22, no. 10, pp. 1345–1359, Oct. 2010.
chine learning based techniques,” in Proc. IEEE 17th India Council Int. [27] P. Qin and L. Zhao, “A novel transfer learning-based cell SOC online es-
Conf., 2020, pp. 1–6. timation method for a battery pack in complex application conditions,”
[5] S. Hou, Y. Chu, and J. Fei, “Intelligent global sliding mode control using IEEE Trans. Ind. Electron., vol. 71, no. 2, pp. 1606–1615, Feb. 2024.
recurrent feature selection neural network for active power filter,” IEEE [28] S. Zhang, H. Zhu, J. Wu, and Z. Chen, “Voltage relaxation-based
Trans. Ind. Electron., vol. 68, no. 8, pp. 7320–7329, Aug. 2021. state-of-health estimation of lithium-ion batteries using convolutional
[6] I. Martin-Diaz, D. Morinigo-Sotelo, O. Duque-Perez, and R. J. Romero- neural networks and transfer learning,” J. Energy Storage, vol. 73, 2023,
Troncoso, “An experimental comparative evaluation of machine learn- Art. no. 108579.
ing techniques for motor fault diagnosis under various operating [29] M. Zhang, X. Wang, D. Yang, Z. Cui, and M. G. Christensen,
conditions,” IEEE Trans. Ind. Appl., vol. 54, no. 3, pp. 2215–2224, “Transfer learning for identifying impedance estimation in voltage
May/Jun. 2018. source inverters,” in Proc. IEEE Energy Convers. Congr. Expo., 2020,
[7] Y. León-Ruiz, M. González-García, R. Alvarez-Salas, J. Cuevas- pp. 6170–6174.
Tello, and V. Cárdenas, “Fault diagnosis based on machine learning [30] M. Zhang, Y. Zhang, and Q. Xu, “Transfer learning based online
for the high frequency link of a grid-tied photovoltaic converter impedance identification for modular multilevel converters,” IEEE
for a wide range of irradiance conditions,” IEEE Access, vol. 9, Trans. Power Electron., vol. 38, no. 10, pp. 12207–12218, Oct. 2023.
pp. 151209–151220, 2021. [31] Y. Wang and T. Wu, “Temperature rising model of tubular permanent
[8] Z. Li, Y. Gao, X. Zhang, B. Wang, and H. Ma, “A model-data- magnet linear synchronous motor based on transfer learning-deep neu-
hybrid-driven diagnosis method for open-switch faults in power con- ral network,” in Proc. 42nd Chin. Control Conf., 2023, pp. 6707–6712.
verters,” IEEE Trans. Power Electron., vol. 36, no. 5, pp. 4965–4970, [32] J. Liu, T. Wei, N. Chen, W. Liu, J. Wu, and P. Xiao, “A neural network
May 2021. PID controller with dynamic structure adjustment for DC-DC switching
[9] S. Chen, “Review on supervised and unsupervised learning techniques converters,” in Proc. IEEE 7th Int. Conf. Integr. Circuits Microsystems,
for electrical power systems: Algorithms and applications,” IEEJ Trans. 2022, pp. 356–360.
Elect. Electron. Eng., vol. 16, no. 11, pp. 1487–1499, 2021. [33] M. Jafari and Z. Malekjamshidi, “Design, simulation and implementa-
[10] E. Mohammadi, M. Alizadeh, M. Asgarimoghaddam, X. Wang, and M. tion of an adaptive controller on base of artificial neural networks for a
G. Simões, “A review on application of artificial intelligence techniques resonant dc-dc converter,” in Proc. IEEE 9th Int. Conf. Power Electron.
in microgrids,” IEEE J. Emerg. Sel. Top. Ind. Electron., vol. 3, no. 4, Drive Syst., 2011, pp. 1043–1046.
pp. 878–890, Oct. 2022. [34] H. Maruta and D. Hoshino, “Transient response improvement of
[11] E. O. Arwa and K. A. Folly, “Reinforcement learning techniques for repetitive-trained neural network controlled DC-DC converter with
optimal power control in grid-connected microgrids: A comprehensive overcompensation suppression,” in Proc. 45th Annu. Conf. IEEE Ind.
review,” IEEE Access, vol. 8, pp. 208992–209007, 2020. Electron. Soc., 2019, pp. 2088–2093.
[12] D. Cao et al., “Reinforcement learning and its applications in modern [35] R. Wang, Z. Zhou, and G. Qu, “Fuzzy neural network PID control based
power and energy systems: A review,” J. Modern Power Syst. Clean on RBF neural network for variable configuration spacecraft,” in Proc.
Energy, vol. 8, no. 6, pp. 1029–1042, 2020. IEEE 3rd Adv. Inf. Technol. Electron. Automat. Control Conf., 2018,
[13] Y. Gao, T. Yang, S. Bozhko, P. Wheeler, and T. Dragičević, “Fil- pp. 1203–1207.
ter design and optimization of electromechanical actuation systems [36] M. P. Belov, D. D. Truong, and P. Van Tuan, “Self-tuning PID con-
using search and surrogate algorithms for more-electric aircraft applica- troller using a neural network for nonlinear exoskeleton system,”
tions,” IEEE Trans. Transport. Electrific., vol. 6, no. 4, pp. 1434–1447, in Proc. IEEE 2nd Int. Conf. Neural Netw. Neurotechnol., 2021,
Dec. 2020. pp. 6–9.

374 VOLUME 4, 2023

[37] Z. Li, J. Pang, C. Xu, and C. Zhang, “Electric vehicle charging system [59] S. Hou, Y. Chu, and J. Fei, “Robust intelligent control for a class of
pressure control based on fuzzy neural network pid control,” in Proc. power-electronic converters using neuro-fuzzy learning mechanism,”
IEEE Conf. Telecommun. Opti. Comput. Sci., 2021, pp. 873–877. IEEE Trans. Power Electron., vol. 36, no. 8, pp. 9441–9452, Aug. 2021.
[38] D. Xu, J. Liu, X.-G. Yan, and W. Yan, “A novel adaptive neural network [60] M. Demirtaş, I. Cetinbas, S. Serefoğlu, and O. Kaplan, “Ann controlled
constrained control for a multi-area interconnected power system with single phase inverter for solar energy systems,” in Proc. 16th Int. Power
hybrid energy storage,” IEEE Trans. Ind. Electron., vol. 65, no. 8, Electron. Motion Control Conf. Expo., 2014, pp. 768–772.
pp. 6625–6634, Aug. 2018. [61] A. Oshaba, E. S. Ali, and S. M.Abd Elazim, “Pi controller design
[39] M. K. Das, P. Bera, P. Sarkar, and K. Chakrabarty, “PI-RLNN controller using artificial bee colony algorithm for MPPT of photovoltaic system
for LFC of hybrid deregulated power system based on SPOA,” in Proc. supplied DC motor-pump load,” Complexity, vol. 21, no. 6, pp. 99–111,
IEEE 18th India Council Int. Conf., 2021, pp. 1–5. 2016.
[40] K. Rouzbehi, A. Miranian, E. Rakhshani, A. Luna, and P. Rodriguez, [62] L. Galotto, J. Pinto, J. Bottura Filho, and G. Lambert-Torres, “Recur-
“Identification and local linear control of a DC-DC buck converter using sive least square and genetic algorithm based tool for PID controllers
local model networks,” in Proc. Int. Conf. Renewable Energy Res. Appl., tuning,” in Proc. IEEE Int. Conf. Intell. Syst. Appl. Power Syst., 2007,
2013, pp. 745–750. pp. 1–6.
[41] J. Liu, T. Wei, N. Chen, J. Wu, and P. Xiao, “Fuzzy logic PID [63] A. Debnath, T. O. Olowu, S. Roy, I. Parvez, and A. Sarwat, “Particle
controller with both coefficient and error modifications for digitally- swarm optimization-based PID controller design for DC-DC buck con-
controlled DC–DC switching converters,” J. Elect. Eng. Technol., verter,” in Proc. North Amer. Power Symp., 2021, pp. 1–6.
vol. 18, pp. 2859–2870, 2023. [64] H. S. Krishnamoorthy and T. N. Aayer, “Machine learning based mod-
[42] S. Vazquez, C. Montero, C. Bordons, and L. G. Franquelo, “Model eling of power electronic converters,” in Proc. IEEE Energy Convers.
predictive control of a VSI with long prediction horizon,” in Proc. IEEE Congr. Expo., 2019, pp. 666–672.
Int. Symp. Ind. Electron., 2011, pp. 1805–1810. [65] G. N. Baltas, N. B. Lai, L. Marin, A. Tarraso, and P. Rodriguez, “Grid-
[43] T. Dragičević, S. Vazquez, and P. Wheeler, “Advanced control methods forming power converters tuned through artificial intelligence to damp
for power converters in DG systems and microgrids,” IEEE Trans. Ind. subsynchronous interactions in electrical grids,” IEEE Access, vol. 8,
Electron., vol. 68, no. 7, pp. 5847–5862, Jul. 2021. pp. 93369–93379, 2020.
[44] J. Rodriguez et al., “State of the art of finite control set model predictive [66] Q. Xu, T. Dragicevic, L. Xie, and F. Blaabjerg, “Artificial intelligence-
control in power electronics,” IEEE Trans. Ind. Informat., vol. 9, no. 2, based control design for reliable virtual synchronous generators,” IEEE
pp. 1003–1016, May 2013. Trans. Power Electron., vol. 36, no. 8, pp. 9453–9464, Aug. 2021.
[45] T. Dragičević and M. Novak, “Weighting factor design in model predic- [67] W. Liu, T. Kerekes, T. Dragicevic, and R. Teodorescu, “Review of
tive control of power electronic converters: An artificial neural network grid stability assessment based on AI and a new concept of converter-
approach,” IEEE Trans. Ind. Electron., vol. 66, no. 11, pp. 8870–8880, dominated power system state of stability assessment,” IEEE J. Emerg.
Nov. 2019. Sel. Topics Ind. Electron., vol. 4, no. 3, pp. 928–938, Jul. 2023.
[46] S. Vazquez et al., “An artificial intelligence approach for real-time [68] J. Ahmed and Z. Salam, “An improved perturb and observe (P&O) max-
tuning of weighting factors in FCS-MPC for power converters,” IEEE imum power point tracking (MPPT) algorithm for higher efficiency,”
Trans. Ind. Electron., vol. 69, no. 12, pp. 11987–11998, Dec. 2022. Appl. Energy, vol. 150, pp. 97–108, 2015.
[47] Y. Gao et al., “Inverse application of artificial intelligence for the control [69] D. Sera, L. Mathe, T. Kerekes, S. V. Spataru, and R. Teodorescu, “On
of power converters,” IEEE Trans. Power Electron., vol. 38, no. 2, the perturb-and-observe and incremental conductance MPPT methods
pp. 1535–1548, Feb. 2023. for PV systems,” IEEE J. Photovolt., vol. 3, no. 3, pp. 1070–1078,
[48] H. Hussaini, T. Yang, Y. Gao, C. Wang, T. Dragicevic, and S. Bozhko, Jul. 2013.
“Droop coefficient design in droop control of power converters for [70] K. Y. Yap, C. R. Sarimuthu, and J. M.-Y. Lim, “Artificial intelligence
improved load sharing: An artificial neural network approach,” in Proc. based MPPT techniques for solar power system: A review,” J. Modern
IEEE 30th Int. Symp. Ind. Electron., 2021, pp. 1–6. Power Syst. Clean Energy, vol. 8, no. 6, pp. 1043–1059, 2020.
[49] D. Wang et al., “A deep neural network based predictive control strat- [71] S. R. Kiran, C. H. Basha, V. P. Singh, C. Dhanamjayulu, B. R. Prusty,
egy for high frequency multilevel converters,” in Proc. IEEE Energy and B. Khan, “Reduced simulative performance analysis of variable
Convers. Congr. Expo., 2018, pp. 2988–2992. step size ann based MPPT techniques for partially shaded solar PV
[50] I. S. Mohamed, S. Rovetta, T. D. Do, T. Dragicević, and A. A. Z. Diab, systems,” IEEE Access, vol. 10, pp. 48875–48889, 2022.
“A neural-network-based model predictive control of three-phase in- [72] S. Muyeen, H. M. Hasanien, and J. Tamura, “Reduction of frequency
verter with an output lc filter,” IEEE Access, vol. 7, pp. 124737–124749, fluctuation for wind farm connected power systems by an adaptive artifi-
2019. cial neural network controlled energy capacitor system,” IET Renewable
[51] M. Novak and T. Dragicevic, “Supervised imitation learning of finite-set Power Gener., vol. 6, no. 4, pp. 226–235, 2012.
model predictive control systems for power electronics,” IEEE Trans. [73] A. Bouzekri, T. Allaoui, M. Denai, and Y. Mihoub, “Artificial
Ind. Electron., vol. 68, no. 2, pp. 1717–1723, Feb. 2021. intelligence-based fault tolerant control strategy in wind turbine sys-
[52] M. Novak and F. Blaabjerg, “Supervised imitation learning of FS-MPC tems,” Int. J. Renewable Energy Res., vol. 7, no. 2, 2017.
algorithm for multilevel converters,” in Proc. 23rd Eur. Conf. Power [74] S. Wang, T. Dragicevic, Y. Gao, S. K. Chaudhary, and R. Teodor-
Electron. Appl., 2021, pp. 1–10. escu, “Machine learning based operating region extension of modular
[53] D. Wang et al., “Model predictive control using artificial neural net- multilevel converters under unbalanced grid faults,” IEEE Trans. Ind.
work for power converters,” IEEE Trans. Ind. Electron., vol. 69, no. 4, Electron., vol. 68, no. 5, pp. 4554–4560, May 2021.
pp. 3689–3699, Apr. 2022. [75] S. Koganti, K. J. Koganti, and S. R. Salkuti, “Design of multi-objective-
[54] S. Wang, T. Dragicevic, Y. Gao, and R. Teodorescu, “Neural network based artificial intelligence controller for wind/battery-connected shunt
based model predictive controllers for modular multilevel converters,” active power filter,” Algorithms, vol. 15, no. 8, 2022, Art. no. 256.
IEEE Trans. Energy Convers., vol. 36, no. 2, pp. 1562–1571, Jun. 2021. [76] A. N. Abdalla et al., “Integration of energy storage system and renew-
[55] S. Wang, T. Dragicevic, G. F. Gontijo, S. K. Chaudhary, and R. Teodor- able energy sources based on artificial intelligence: An overview,” J.
escu, “Machine learning emulation of model predictive control for Energy Storage, vol. 40, 2021, Art. no. 102811.
modular multilevel converters,” IEEE Trans. Ind. Electron., vol. 68, [77] I. Hammoud, S. Hentzelt, T. Oehlschlaegel, and R. Kennel, “Long-
no. 11, pp. 11628–11634, Nov. 2021. horizon direct model predictive control based on neural networks for
[56] W. Zhao, K. Li, and G. W. Irwin, “A new gradient descent approach for electrical drives,” in Proc. 46th Annu. Conf. IEEE Ind. Electron. Soc.,
local learning of fuzzy neural models,” IEEE Trans. Fuzzy Syst., vol. 21, 2020, pp. 3057–3064.
no. 1, pp. 30–44, Feb. 2013. [78] R. Anugula and S. P. K. Karri, “Deep reinforcement learning based
[57] M. R. Mashinchi and A. Selamat, “An improvement on genetic-based adaptive controller of dc electric drive for reduced torque and current
learning method for fuzzy artificial neural networks,” Appl. Soft Com- ripples,” in Proc. IEEE Int. Conf. Technol. Res. Innov. Betterment Soc.,
put., vol. 9, no. 4, pp. 1208–1216, 2009. 2021, pp. 1–6.
[58] R. Hou, D. Shi, Q. Gao, and Y. Hou, “Implementation of adaptive [79] N. Nikdel, P. Nikdel, M. A. Badamchizadeh, and I. Hassanzadeh,
self-organizing fuzzy wavelet neural network: Modeling ac servo sys- “Using neural network model predictive control for controlling shape
tem,” in Proc. IEEE 10th Data Driven Control Learn. Syst. Conf., 2021, memory alloy-based manipulator,” IEEE Trans. Ind. Electron., vol. 61,
pp. 1128–1134. no. 3, pp. 1394–1401, Mar. 2014.

VOLUME 4, 2023 375