IEEE Electrification - December 2022 - 54

500
400
300
200
100
through the RNN cells suffers from vanishing gradients,
especially for long sequences. For example, at the nth element,
the learning from the first element would have
been reduced n - 1 times.
LSTM and gated recurrent unit (GRU) modules were
Net Load
Solar
Building Usage
Figure 3. The net load on a sunny day with high solar irradiance.
BTM solar is able to supplement almost half of the building load.
This decreases the net load requested from the grid.
developed to address vanishing gradients. Both models
use gates with trainable weights that decide which information
passes forward and which is excluded. An LSTM
element controls the flow of information through an
input gate, a forget gate that determines how much
information to keep, and an output gate. The GRU uses
only an update gate, which determines the usefulness of
past information, and a reset gate, which determines
how much of the past information to forget. The gates
can be trained to decide which information to pass to the
output (see Figure 5).
Another development that has proven very useful is
500
400
300
200
100
Net Load
Solar
Building Usage
Figure 4. The net load on a cloudy day with low solar irradiance.
The net load seen at the meter is higher than that in the sunny day
example. Utilities would like to plan for low solar production days
and the increased net load consumption.
CNNs. CNNs are artificial neural networks that are widely
used in image recognition and classification. They are
modeled after how we have trained our brains to precisely
differentiate between objects. For example, we have
trained our brains to accurately distinguish between a
cougar and a leopard by identifying the relevant distinguishing
features-leopards have spots, while cougars do
not. Similarly, CNN models simplify the information by
performing feature extraction, while LSTM provides the
accurate temporal characterization.
The combination has been very effective in forecasting
load but especially the generation of renewable resources.
Techniques for feature extraction like CNN are useful for
determining areas covered by rooftop solar panels and
tracking the movement of clouds affecting solar generation
from images taken by sky cameras or satellite imaging
(see Figure 6).
YY YY Y
H
X
ANN
X
Y
H
Input Layer
Hidden Layer
Output Layer
HHt
Ht+1
Ht+2
XX XX
RNN
YY Y
Long Term
(Cell State)
Short Term
(Previous
Output)
Ht
Ht+1
Ht+2
XX X
LSTM
Figure 5. An overview of an artificial neural network (ANN), RNN, and LSTM. ANN is the foundation
for more sophisticated neural nets like RNN, and LSTM is a variation of RNN with feedback
connections capable of processing sequences of data and images.
54
IEEE Electrification Magazine / DECEMBER 2022
Additional Analytical Use Cases:
Forecasting and Optimization
Many other use cases for analytical
models exist for BTM DERs. Forecasting
and optimization are two
analytical techniques that have
direct applications to BTM devices.
An overview of the use cases is
given in Figure 7. Determining the
future values of renewable generation,
building loads, and sell-back
prices for surplus onsite energy as
well as the charging requirements
of an EV fleet are interesting applications
for forecasting.
Optimization is widely used in
the utility industry for control, economic
operations, and planning. An
abundance of commercial nonlinear,
linear, and mixed-integer optimization
solvers help the industry
8 a.m.
9 a.m.
10 a.m.
11 a.m.
12 p.m.
1 p.m.
2 p.m.
3 p.m.
4 p.m.
5 p.m.
6 p.m.
7 p.m.
8 p.m.
8 a.m.
9 a.m.
10 a.m.
11 a.m.
12 p.m.
1 p.m.
2 p.m.
3 p.m.
4 p.m.
5 p.m.
6 p.m.
7 p.m.
8 p.m.

IEEE Electrification - December 2022

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