IEEE Power & Energy Magazine - May/June 2022 - 17

the Dartmouth Summer Research Project on Artificial Intelligence,
widely recognized as the birth of AI as a research field.
Many theoretical breakthroughs and research advances
were made in the 1950s and 1960s. However, the field of
AI quickly experienced its first winter after Marvin Minsky
and Seymour Papert discovered two drawbacks of ANNs: 1)
single-layer NNs were incapable of processing the exclusiveor
circuit; and 2) computers were not sophisticated enough to
effectively handle the long run time required by large ANNs.
The findings were published in Perceptrons: An Introduction
to Computational Geometry.
The field of AI faced even more challenges in the early
1970s. For example, people realized that many important
applications of AI, such as vision and natural language processes,
required too much information about the world to be
handled by a database at that time. Although some of these
challenges were overcome in later years, they slowed down
the research progress seriously enough for funding agencies
to end financial support.
The second AI wave started in the mid-1970s, when Paul
Werbos proposed a backpropagation algorithm to practically
train multilayer ANNs. In the 1980s, advances in digital
electronics and distributed computing enabled the use of
even larger networks. AI saw its first commercial bubble,
with hundreds of AI companies being incorporated and millions
of dollars flooding into AI research. The bubble, however,
burst rather quickly in the late 1980s and early 1990s
because commercial vendors failed to develop a wide variety
of practical solutions. That was widely considered the second
AI winter. Figure 1(b) summarizes the important years
for AI before the 21st century.
Applications of AI to Load Forecasting
Despite the second winter AI was facing, power engineers
found several AI applications in power systems. In the 1990s,
AI became a hot topic among the power engineering community.
At power engineering society general meetings, when
people asked, " How do we solve 'XYZ' problem? " where
100
120
140
160
20
40
60
80
XYZ might refer to any practical problem such as power quality,
security assessment, fault diagnosis, or load forecasting,
the answer was often as simple as " Artificial intelligence. "
One of the first papers discussing ANNs for load forecasting
was presented at the American Power Conference
in 1989. Since then, thousands of papers have adopted AI
and machine learning (ML) techniques for load forecasting.
The topics included fuzzy expert systems, support vector
machines, and deep learning. Fuzzy expert systems are
based on fuzzy logic, which models logical reasoning with
vague or imprecise statements rather than straightforward
true-or-false statements. Support vector machines are models
that can find a hyperplane in a high-dimensional space
to classify observations into one category or the other. Deep
learning is based on ANN, where the word deep refers to the
use of " many " hidden layers in the network.
Figure 3 depicts the number of journal articles through
2021 with " load forecasting " in the title and selected AI/
Input Layer
Hidden Layer
Output Layer
figure 2. A simple three-layer, feedforward ANN.
Neural Network
Fuzzy
Support Vector
Deep
ALL
100
150
200
250
50
figure 3. The journal papers with " load forecasting " in the title and " AI/ML techniques " in the abstract.
may/june 2022
ieee power & energy magazine
17
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021

IEEE Power & Energy Magazine - May/June 2022

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