IEEE Power & Energy Magazine - May/June 2022 - 22

proposed approach at an unfair advantage. Meanwhile, readers
who do not know these details can hardly make the proposed
technique work well in real-world data.
Clarification 5
The accuracy values reported by many academic papers
are too good to be true. Some reported values have been
significantly exaggerated due to the flawed process of
model building and forecast evaluation, such as peeking
and cherry-picking. Figure 6 lists a few checkpoints for
readers to conduct a smoke test when they see astonishingly
good results.
Opportunities
Many opportunities are presented for today's researchers
and practitioners interested in load forecasting. Modern
computers are more powerful than ever before. In the 1980s,
computing resources were a major concern when people
developed load forecasting models. Many models had to be
built offline and retrained every week or month. The methods
at that time had to find a tradeoff between using the most
recent data and how fast the models could be estimated. Just
like computer technologies pushed AI to its second hype in
the 1980s, today's computing infrastructure enables applications
of deep learning and other computationally intensive
techniques to load forecasting.
In Dryar's 1944 paper, the weather was first identified as
a driving factor of electricity demand. In the 20th century,
the data used by load forecasters were quite limited. Weather
data came from a few weather stations in the service territory.
Other than some interval meters collecting hourly
demand for selected households for load research, there were
no advanced communicating (smart) meters. Utilities had
only monthly or bimonthly meter readings for most customers.
For long-term load forecasting, monthly or annual load
data were used.
Today's load forecasters can access a variety of data
sources and high granularity. Smart meters are collecting
load readings every hour, if not more frequently. Weather
conditions are monitored by thousands of in situ weather stations
and satellites. In many places, power companies have
installed cameras to capture sky images. Meanwhile, demographics
and economic information are monitored and collected
in great detail. Social media platforms also generate
useful data for load forecasters to better understand human
behaviors and local environments.
Forecasting competitions can significantly stimulate
the advance of a forecasting field. A well-organized competition
can recognize state-of-the-art methods, attract talents
from other fields to solve a problem of interest, and
release benchmark data for future researchers to continue
the investigation.
Table 2 lists five major load competitions during the
figure 6. The smoke tests for too-good-to-be-true results.
last three decades. Before the 2010s, load forecasting
literature saw only two remarkable competitions. One was
supported by Puget Sound Power and Light Company in
1991-1992, and the other was organized by the EUNITE
Time
Competition Name
2001
2012
2014
2017
table 2. Five notable load forecasting competitions from the 1990s to 2010s.
Problem
1991-1992 Puget Sound Shootout
GEFCom2012
GEFCom2014
GEFCom2017
EDF: Électricité de France.
22
ieee power & energy magazine
may/june 2022
Short-term load forecasting
EUNITE Competition Medium-term load forecasting
Hierarchical load forecasting
Probabilistic load forecasting
Affiliations of Winning Teams
University of California, San Diego
Southern Methodist University
National Taiwan University
CountingLab
Cambridge University
EDF R&D
EDF R&D
University of North Carolina at Charlotte
Hierarchical probabilistic load forecasting Japan Meteorological Corporation
* The paper is proposing a mix (or hybrid) of
many techniques.
* The paper is merely catching new buzzwords.
* The data is proprietary.
* The paper is not co-authored with industry
people (or not sponsored by the industry).
* The proposed method does not utilize new
variables
* The proposed method does not take knowledge
of other domains.
* The proposed method does not leverage
new computing resources.
* The proposed method is dominating its
counterparts (another credible method)
in all aspects.
IEEE Power & Energy Magazine - May/June 2022

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