IEEE Electrification - December 2022 - 53

multiple supervised learning models can provide better
model accuracy. Supervised learning is generally more
accurate and reliable, as the inputs are well known, with
labels identifying the characteristics and properties of
the data.
Unsupervised models are models that can process data
where the outcome is unknown. These models are commonly
clustering models, association rules like market
basket analysis, and variable reduction techniques such as
principal component analysis. Unsupervised models are
frequently used in customer analytics, such as customer
segmentation. Principal component analysis is used as a
preprocessing step to predictive modeling to decrease the
number of redundant input variables and even combine
highly correlated variables into new uncorrelated features.
Unlike supervised learning, unsupervised models are useful
in cases where data are not labeled and inputs are
unidentified. There are fewer examples of unsupervised
learning models used in the utilities industry and most
notable as a variable reduction method.
Reinforcement learning, a core machine learning paradigm,
focuses on training models by rewarding desirable
behaviors and punishing undesired ones. Rewarding
a particular behavior strengthens the frequency of the
behavior, whereas penalizing an undesirable behavior is
accomplished by strengthening a behavior that leads to
the avoidance of negative conditions. Recently, reinforcement
learning was used on SAS Institute's campus
to optimize heating, ventilation, and air conditioning
operations by using the minimization of energy costs as
the reward function while maintaining occupant comfort.
Reinforcement learning techniques can also be
applied to data centers as well to minimize energy usage
while maintaining the necessary temperature and
humidity requirements.
Reinforcement learning traditionally requires many
samples and " exploration " stages to develop solutions and
policies. The ability to expedite and reduce the need for
exploration can be addressed with transfer learning,
which is being used, for example, in the study of DERs
across multiple geographic locations. Transfer learning
relies on knowledge collected in solving similar previous
tasks to bias the learning process on a new task. This can
result in much more efficient exploration during the
learning process of the model.
Transfer learning is a method of machine learning
where a model developed and trained for one use is
reused as the starting point for another similar task. The
purpose is to have a model gain knowledge learned from
one task for the sake of a second. The value in wielding
knowledge in a model from one task to another is the
potential of efficient reinforcement learning.
There is ongoing experimentation using transfer learning
to develop models for forecasting net load. The net
load is the total electric demand minus the variable generation.
This approach takes models trained on data in one
operating region and reuses them in another operating
region. The benefit of this type of model is that it is more
efficient, requiring fewer data to produce similar results.
The other benefit is that regions with rich sensor data can
train models that are then transferred and used in regions
where only substation data are available.
Knowledge is transferred in the form of an established
correlation between solar irradiance and distributed photovoltaic
(PV) generation. With this training and knowledge,
the new region that lacks BTM data, having only net
load measurements, can properly produce a net load forecast.
The reuse of a model trained in the data-rich area
prevents the need for extensive retraining in the new
operating region and leads to more accurate forecasts.
Forecasting the Net Load: An Application
for BTM DERs
Forecasting the net load is a critical task for utilities, but it
can be complicated by BTM resources. Renewable energy
sources, such as solar and wind, provide electricity when
weather conditions are good but place these consumers
back onto the grid when conditions are unfavorable.
Because the energy produced is used before the meter,
utilities have no insight into real-time onsite production
and total consumption. They can only see the net load at
the meter. This makes it difficult for utilities to plan ahead
and purchase energy. Figures 3 and 4 illustrate the varied
effects of weather on solar generation and the resulting
change in consumption from the grid (net load).
As mentioned, we can develop and deploy a supervised
learning model that is trained on a sample of known
solar/wind production data to quantify the relationship
between weather conditions and energy production. BTM
generation can then be estimated for a wider area by
using the weather-based model in conjunction with the
net load profile over time. This enables the ability to more
accurately predict weather-dependent BTM DER production
at a site- or region-specific level. In addition to this, a
forecasting model built for one region can be refit and
used in another region as well. Analytical software
enables the easy and automated refitting of a model to a
different region, which may have a different climate than
the original location.
When DERs are more accurately predicted at highinterval
time granularities, this allows for better planning
capabilities-especially when fossil fuel backups are necessary
to backfill DER availability. Models, such as RNNs,
convolutional neural networks (CNNs), random forest
models, and time series models are being applied today to
solve these challenges.
RNNs are a subset of neural networks that are specifically
designed for forecasting time series data, such as
BTM production and loads. Each element in a sequence is
a neural network that performs the same task. Aside
from its inputs, each element has a memory of what happened
in the past. However, information traveling
IEEE Electrification Magazine / DECEMBER 2022
53
IEEE Electrification - December 2022

Table of Contents for the Digital Edition of IEEE Electrification - December 2022
