IEEE Electrification - December 2022 - 91

x " A Gallium Nitride-Based Superresolution
Method for Generating
High-Resolution Load Profiles, "
by Yiyan Li, North Carolina State
University, USA
x " OpenGridGym: An AI-Friendly
Simulation and Analytics Platform
for Power Distribution Systems, "
by Le Xie and Rayan
El-Helou, Texas A&M University,
USA.
Supersession " Artificial
Intelligence in Power Systems "
at the PES GM 2022
x " Review and Vision of AI in Power
System Applications, " by Fran Li,
University of Tennessee, USA
x " Machine Learning for Power
Systems: From Pure Data-Driven
to Physics-Informed Methods, " by
Nanpeng Yu, University of California
at Riverside, USA
x " AI for Grid Operation and Emergency
Control: Progress and Perspectives, "
by Qiuhua Huang and
Yousu Chen, Pacific Northwest
National Laboratory, USA
x " AI-Based Online Fast Stability
Assessment Covering a Large
Dispatch Range, " by Yilu Liu, University
of Tennessee and Oak
Ridge National Laboratory, USA
x " Machine Learning for Security
Assessments in a Low-Inertia
Grid, " by Jochen Cremer, Delft
University of Technology, The
Netherlands
x " The Role of Information Models
and Digital Twins to Support AI
Applications for the Power Grid, "
by Mark McGranaghan, Electric
Power Research Institute, USA
x " Physics-Inspired ML for Power
Grid Data Anomalies and Events
Detection, " by Anurag Srivastava,
West Virginia University, USA.
Successes of ML in Power
Why has ML been very successful in
recent years? The success of this new
wave of ML in the last decade can be
largely attributed to the advancement
of deep learning (DL)
techniques, such as deep neural networks
(DNNs), as well as the hardware
development in GPUs and big datasets.
In comparison with a traditional
artificial neural network (ANN), which
usually needs a human expert to
extract high-level features from raw
data and then apply these features as
the input to the ANN, DL technologies
can extract high-level features on
their own from raw data, so they may
take more time for training but will
achieve significant acceleration during
testing and field applications.
In the power systems domain, the
success of ML can be classified into
three categories. First, ML can be
applied to solve computationally
challenging problems. Examples may
include multiple-scenario security
screening, transient stability assessment,
complex or nonconvex optimization
or dispatch problems involving
multiple levels or multiple stages,
fault and abnormality detections, etc.
In such computationally intractable
tasks, a well-trained DNN, possibly
combined with other techniques, like
a convolutional neural network, can
achieve an extremely fast solution in
regression or classification once the
network is well trained, in comparison
with traditional model-based
approaches. Various studies reported
hundreds or even thousands of times
the computational acceleration. This
is not surprising at all if one is familiar
with the achievements of DL in
other domains.
Another category of ML applications
in power is for power grid control
problems involving a Markov
decision process, which use reinforcement
learning (RL) as an " intelligent "
trial-and-error approach. Usually,
RL is combined with DL to form
a deep RL (DRL) approach to achieve
a model-free, data-driven solution.
Many present control room functions
that are based on lookup tables from
offline training could be potentially
replaced by such DRL-based solutions,
which may provide a much
more accurate, fast, and dynamic
control action. Also, some control
problems that are difficult to model
with a set of physical equations (e.g.,
the behavior of consumers in a
demand-response program) have
been reportedly well solved with DRL
approaches in field demonstrations.
The third category is for data processing.
ML techniques have a strong
ability to learn statistical features
hidden in a dataset. In a completely
data-driven way, unsupervised and
" label-free " ML techniques automatically
capture data features. This
avoids mathematics- or experiencebased
manual data feature extraction,
which is difficult to do when the
dataset has high-dimensional information.
ML-supported data processing
can be independently applied to
handle data collected from power
grids or renewable power plants. It
also can be combined with other DL/
DRL methods since it automatically
refines the original database via filtering
or extension for further ML
training. Currently, ML techniques
have shown promising performances
in data forecasting, bad data identification,
scarce data augmentation,
and so on.
Challenges of ML in Power
The success of ML applications is
mostly due to the superior performance
of DL, which may transit
many offline rule-based operations
to true real-time operations. However,
there are a few challenges that
must be well addressed for ML to be
broadly applied by the power industry.
First, the performance guarantee
of ML must be achieved. The power
industry is different from other
areas due to its critical requirements
of security and safety. A failure of
ML approaches in a control room
can immediately cause a catastrophic
event, especially if this
involves security functions, so a
guarantee of performance is needed.
Furthermore, missing an upcoming
(continued on page 96)
IEEE Electrification Magazine / DECEMBER 2022
91
IEEE Electrification - December 2022

Table of Contents for the Digital Edition of IEEE Electrification - December 2022
