IEEE Power & Energy Magazine - November/December 2021 - 85

fronts; patchy, fast-moving clouds; and gusting wind conditions.
It is generally difficult to accurately forecast electricity
demand as well as wind and solar generation for these
conditions, and this leads to higher uncertainty in the forecasts
feeding into the process.
Recognizing the increasing importance of understanding
forecast uncertainty to the operation of a secure power
system, in 2018, a change to the rules governing Australia's
electricity system introduced a probabilistic assessment of
the forecast inputs into the week-ahead resource assessment
process. To account for the uncertainty inherent in each of
the forecast inputs, the Australian Energy Market Operator
developed a machine learning model (based on a Bayesian
belief network) trained on more than 20 million data points
of historic forecast errors.
Using the forecast weather conditions for the week ahead
as input, the machine learning model gives a conditional
probability of the expected level of forecast error under those
conditions. The expected forecast error from the network is
then used in the week-ahead resource assessment to indicate
the amount of reserve generation above the predicted level of
demand necessary to ensure a reliable energy supply.
Innovation and Future Research
The accuracy of operational renewable energy forecasting
tools remains low in challenging situations such as evolving
weather fronts. Improved accuracy translates into higher
competitiveness of RESs in markets and more economically
efficient and safe operation of the power system.
These factors stimulate intense research internationally
in the field of renewable energy forecasting. Considering
the model chain and associated gaps and bottlenecks
shown in Figure 1, numerous directions of future research
emerge. Some stem from new sources of available data (i.e.,
lidars, radars, sky cameras, and satellites), allowing better
modeling of weather variables and RES production. NWP
models may benefit from RES plants as distributed sensors
for data assimilation.
RES forecasting models need to consider multiple data
sources, in contrast to the mainstream approach, where tailored
models are built upon for specific data sources (i.e.,
satellite images for PV predictions up to 6 h ahead). More
general approaches may lead to a simpler model chain where
forecasting models can cover multiple timeframes.
Beyond convergence in the temporal scales, models based
on hierarchical forecasting are needed to enable the compatibility
of predictions and associated uncertainty at different
geographical scales. Recent work has shown that using data
from neighboring RES sites improves the accuracy of a site.
One of the related directions is to develop approaches for
data sharing that respect privacy and confidentiality constraints,
accompanied by data market concepts to facilitate
and incentivize data owners to share their data.
Contributions from artificial intelligence techniques may
support massive data flow processing. They can be used to
november/december 2021
develop prescriptive analytics for decisions (i.e., trading)
based on the input data while replacing multiple forecasting
and optimization steps, as shown in Figure 8.
Finally, new application use cases continuously appear
where classical or new forecasting products are required.
The main challenge is how to develop decision support
tools that efficiently account for the inherent uncertainties,
a massive amount of streaming data, and features to support
human-in-the-loop informed decision making. Multidisciplinary
research initiatives are being undertaken that
demonstrate the benefits of collaborative efforts to tackle the
complex nature of these problems.
For Further Reading
B. Li and J. Zhang, " A review on the integration of probabilistic
solar forecasting in power systems, " Solar Energy, vol.
207, no. 10, pp. 777-795, Sept. 2020. doi: 10.1016/j.solener
.2020.06.083.
C. Sweeney, R. J. Bessa, J. Browell, and P. Pinson, " The
future of forecasting for renewable energy, " Wiley Interdisciplinary
Rev., Energy Environ., vol. 9, no. 2, p. e365, Mar./
Apr. 2020. doi: 10.1002/wene.365.
F. Fahiman, S. Disano, S. M. Erfani, P. Mancarella, and
C. Leckie, " Data-driven dynamic probabilistic reserve sizing
based on dynamic Bayesian belief networks, " IEEE
Trans. Power Syst., vol. 34, no. 3, pp. 2281-2291, May 2019.
doi: 10.1109/TPWRS.2018.2884711.
L. Cavalcante, R. J. Bessa, M. Reis, and J. Browell,
" LASSO vector autoregression structures for very shortterm
wind power forecasting, " Wind Energy, vol. 20, no. 4,
pp. 657-675, Apr. 2017. doi: 10.1002/we.2029.
X. G. Agoua, R. Girard, and G. Kariniotakis, " Short-term
spatio-temporal forecasting of photovoltaic power production, "
IEEE Trans. Sustain. Energy, vol. 9, no. 2, pp. 538-
546, 2018. doi: 10.1109/TSTE.2017.2747765.
Biographies
Jack Fox is with the Australian Energy Market Operator,
Melbourne, Australia.
Erik Ela is with the Electric Power Research Institute,
Knoxville, Tennessee, USA.
Ben Hobbs is with Johns Hopkins University, Baltimore,
Maryland, USA.
Justin Sharp is with Sharply Focused, Portland, Oregon,
USA.
Josh Novacheck is with the National Renewable Energy
Laboratory, Denver, Colorado, USA.
Amber Motley is with the California Independent System
Operator, Folsom, California, USA.
Ricardo J. Bessa is with INESC TEC, Lisbon, Portugal.
Pierre Pinson is with the Technical University of Denmark,
Lyngby, Denmark.
Georges Kariniotakis is with ARMINES, Nice,
France.
p&e
ieee power & energy magazine
85
IEEE Power & Energy Magazine - November/December 2021

Table of Contents for the Digital Edition of IEEE Power & Energy Magazine - November/December 2021
