IEEE Power & Energy Magazine - May/June 2022 - 18

ML techniques in the abstract. In chronological order, load
forecasting literature, and journal articles in particular,
found ANNs in 1991, fuzzy expert systems in 1992, support
vector machines in 2002, and deep learning in 2015.
In total, 1,984 journal articles with " load forecasting " in
the title were published from 1991 to 2021, of which 1,245
mentioned AI/ML techniques in the abstract. ANNs are
mentioned in 796 of them. Indeed, ANNs have been the
most popular technique in load forecasting literature during
the past three decades.
Most of these ANN models only stayed in academic
papers or the laboratory environment. Few received much
attention or were accepted by the industry. Nevertheless, one
branch of research was able to make load forecasting one of
the most successful ANN applications. In 1995, IEEE Transactions
on Power Systems published an article describing a
load forecasting system based on ANNs, which was implemented
at 20 United States utilities and used by several of
them. The system was later named an ANN short-term load
forecaster (ANNSTLF). In 1997, IEEE Transactions on Neural
Networks published an article describing ANNSTLF's
second generation, with 32 utilities across the United States
and Canada using the system. A year later, IEEE Transactions
on Power Systems published an article describing the
third generation of ANNSTLFs used by 35 utilities in the
United States and Canada.
Although all three generations of ANNSTLFs are based
on ANNs, their structures are quite different. The first generation
included 38 ANNs and 24 combiners. Each ANN had
a three-layer, feedforward structure, as shown in Figure 2.
Some of them had nine input neurons, while others had 72.
The 38 ANNs were designed to capture load and weather
relationships. The outputs from these ANNs were fed to 24
combiners to generate a forecast for the 24 h of a day. The
second generation eliminated the combiners and reduced the
number of ANNs to 24, one for each hour of the day. The
ANNs were separated into four groups for different periods
of the day. The third generation consisted of two ANN load
forecasters and one combiner. One ANN was trained to predict
the regular (base) load of the next day, while the other
ANN predicted the change in hourly load from yesterday to
today. The final load forecast for each hour was the linear
combination of the outputs from these two ANNs. The commercialized
version of the system later became one of the
major load forecasting solutions in the power industry, used
by many power companies and energy trading firms.
Load forecasting was widely recognized as a successful
application of ANNs in the 1990s, but in the following
decade, the load forecasting community made little progress
in terms of methodological innovation and improvements
on model accuracy and practicality. Meanwhile, other communities,
such as computer vision and mathematical programming,
were making significant progress both methodologically
and practically. Looking back at load forecasting
literature in the 1990s-2000s and comparing it with those
flourishing areas, we can find several reasons for the slow
progress:
✔ The load forecasting literature had no benchmarking
data or models until the 2010s.
✔ Many load forecasting papers were not reproducible.
✔ Many researchers failed to make direct comparisons
with other state-of-the-art load forecasting models.
As a result, load forecasting literature has been filled with
papers of varying degrees of quality. Many have created
illusions among the load forecasting community, further distracting
it from effectively making the next breakthroughs.
Some of the aforementioned issues have been or are in the
process of being resolved, while others may still take some
time to overcome. The remainder of this article will focus on
clearing up these illusions.
Modeling Process
Figure 4 presents a typical load forecasting process. At
first, a load forecaster would gather load and weather history
data to build a load forecasting model. By putting the
model and weather forecast together, the load forecaster
can generate load forecasts. Some load forecasters may
exclude weather data when the load is not weather-sensitive,
the weather data are not reliable, or the modeling techniques
do not require weather data. Different organizations may
Load
History
Model
Modeling
Process
Weather
History
figure 4. A typical load forecasting process.
18
ieee power & energy magazine
may/june 2022
Weather
Forecast
Forecasting
Process
Load
Forecast
IEEE Power & Energy Magazine - May/June 2022

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