IEEE Power & Energy Magazine - May/June 2022 - 15

the construction of the Pearl Street Station in 1882, the first
power plant in the United States, serving an initial load of
400 lamps for 82 customers in lower Manhattan Island. The
load profiles during those early years were as simple as step
functions, while load forecasting was as easy as counting
the light bulbs turned on in the evening. The same method is
still being used today by many local distribution companies
to estimate the load of streetlights.
The first large-scale, alternating current (AC) electric
generating plant in the world, Niagara Falls Power Plant,
was built by Nikola Tesla and George Westinghouse in
1895. The AC systems made long-distance power transmission
an economically viable solution, which helped power
companies realize economies of scale. With larger power
generators, electricity could be produced at a lower unit
cost. Meanwhile, power companies started to encourage
customers to use electricity, further stimulating the invention
of electric appliances. As electricity end uses became
more diverse, people started to find more curvatures in the
load profiles driven by human and business activities, such
as turning lights on and off, using electric irons, listening
to radio broadcasts, and shopping during holiday seasons.
Most of these activities can be captured by calendar variables,
just as in today's load forecasting models.
The first modern air conditioner was invented in 1902 by
Willis Carrier. Air conditioners not only brought comfort to
our daily lives but presented challenges to calendar-based
load forecasting methods. As the market penetration of air
conditioners increased, the effect of weather started to play a
major role in load profiles. In 1944, Henry A. Dryar, a chief
load dispatcher at Philadelphia Electric Company, stated
" there is a variable component of the load that reflects the
effect of the weather. " He described three weather variables:
temperature, wind velocity (wind speed), and cloudiness
(cloud cover). Load forecasting then started to attract attention
from the power industry and the research community.
Figure 1(a) summarizes the important years for load forecasting
before the 1950s.
The Electronic Numerical Integrator and Computer
confusion, exaggerated beliefs, and misleading ideas) of
what AI can do.
Today, most of the AI-based load forecasting models proposed
in academic literature are still at the theoretical level,
with few adopted in practice. In this article, we examine five
illusions associated with developing AI-based models. We
also present solutions and clarifications to help improve the
efficiency of AI for load forecasting, given opportunities in
the big data era.
Prepersonal Computer-Era
Load Forecasting
After inventing the incandescent light bulb in 1879, Thomas
Edison began a series of efforts to commercialize his lightbulb
and demonstrate his electric lighting systems. One was
may/june 2022
(ENIAC), the first programmable, electronic, general-purpose
digital computer was made in 1945. Before the wide
adoption of personal computers (PCs), many methods were
already being applied to load forecasting. Dryar's work represents
some of them, assigning weights to different variables
to forecast the system load. The concept is like multiple
linear regression, which uses several explanatory variables
to predict the outcome of a response variable. Apparently, at
that time, the parameters were not estimated by computers
running statistical software packages.
Some load forecasters kept a book of temperature profiles
matched with load profiles for different day types in different
seasons. In the morning, they looked at the weather forecast,
picked the closest temperature profile from the book,
and found the matching load profile for that day. This is also
known as the similar day method, which is still used in many
ieee power & energy magazine
15
IEEE Power & Energy Magazine - May/June 2022

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