IEEE Power & Energy Magazine - May/June 2022 - 33

Short Term Load Forecaster, a tool developed by the Electric
Power Research Institute in the United States; and a few
approaches have been related to decision trees for security
assessment. The main reason for the slow adoption of ML in
the power industry is that most ML approaches, such as neural
networks and others, are considered a black box.
Who would trust a black box to avoid a blackout or to
find an optimal operating point that will not violate any line
limits? Who would trust a black box for any safety-critical
application, e.g., in energy, health care, or automobiles? The
risk is just too high.
What does it take for ML to get adopted? For insight into
the adoption of novel methods, not only related to ML, consider
the following two aspects:
1) To what extent does the industry already have solutions
to a problem?
2) What is the risk associated with using a novel method
instead of a conventional one?
A novel method will only be seriously considered if it falls
below a risk threshold. If it satisfies the risk requirements, its
better performance, such as higher accuracy or computation
speed, can lead to adoption. Figure 1 presents an illustration
of these considerations.
ML, and artificial intelligence (AI) in general, has been
particularly successful in areas where there are little to no
established solution methods. AlphaGo is an AI program
developed by DeepMind Technologies that harnessed the
complexity of the game Go and repeatedly beat the human
world champion and other software programs. By itself, this
feat is impressive if we consider that Go is an ancient Chinese
game with an astronomical number of possible combinations
of player moves (2.1 × 10170), vastly greater than the
number of known atoms in the known, observable universe
(1 × 1080). AlphaFold is an AI program that predicted, better
than any other program, protein structures that were impossible
to predict for the past 50 years. Both AlphaGo and AlphaFold
yielded results never before possible. Because the risk of
catastrophe associated with their task was nonexistent, their
adoption was extremely rapid. On the other hand, self-driving
cars, which heavily use ML algorithms to detect surroundings,
are also attempting to solve
an unsolved problem, but in several
cases, and despite significant
progress, we are still above the risk
threshold when it comes to their
widespread deployment without
the involvement of a human driver.
Power systems are different.
Risk Level
Unacceptable
Autonomous
Driving
Threshold
We have spent the past 100 years
trying to understand how electricity
flows along power lines and
what happens in the voltage and
current right after a disturbance.
And we have managed that quite
well. The models, despite being
may/june 2022
Acceptable
AlphaGo
AlphaFold
No Methods
Load Forecasting
Smart Charging Agent
Predictive Maintenance of Grid Assets
Established Methods
Maturity
figure 1. The adoption of novel methods (blue area) in light of the existence or nonexistence
of alternative solution techniques, and the risks the novel processes pose.
ieee power & energy magazine
33
ML in Security
Assessment
computationally intensive, can predict well how voltage and
current behave in many operating conditions. But compared
to a few decades ago, power systems have become significantly
more complex. The connection of millions of power
electronic devices leads to rapidly evolving phenomena,
requiring the inclusion of more complex models and a faster
decision-making process. Distributed renewable generation
and electric vehicles add a lot of uncertainty and create
thousands of new injection points that continuously change
the balance between energy supply and demand. This
all means one thing: if a system operator wants to make
sure that no blackout happens (a procedure called security
assessment), he or she needs to run not only more complex
models but also for a lot more scenarios at faster paces than
before. A security assessment that once took a few hours to
carry out would now need a few days, which is unacceptable
for meeting real-time grid-operational requirements.
ML methods, on the other hand, can continuously learn
and adapt to their environment and are extremely fast when
computing an output. Such practices have been shown to
outperform conventional techniques, e.g., predictive maintenance
of transmission lines and transformers, or smart
charging of electric vehicles. ML approaches are gradually
being adopted for such applications because of their
low risk to power system operation. However, handling the
sheer complexity of power system operation procedures
to avoid blackouts (security assessment) or determine an
optimal operating point without violating any single operational
constraint is still too " safety critical " to accept an
ML-based solution. This, despite early promising results
reported in the research literature that ML can help. Power
system operators find it difficult to trust methods they do
not understand, and which have thus far provided no performance
guarantees.
Researchers have recently been working to address
this skepticism and remove barriers, allowing ML to enter
power system applications, exploiting its benefits. In the
first transition period of using ML tools for power system
safety-critical operations, the authors view ML algorithms
as " assistants " to established procedures in the form of a

IEEE Power & Energy Magazine - May/June 2022

Table of Contents for the Digital Edition of IEEE Power & Energy Magazine - May/June 2022