IEEE Electrification - December 2022 - 26

x ADM: This approach uses aggregate models to estimate
BTM resource adoption under influential factors.
It focuses on the behavior of the whole group of
customers. These models are estimated by fitting
them to the historical adoption data through regression
techniques. For example, in the famous Bass diffusion
model, the adoption rate is determined by two
factors: the innovation factor (representing innovative
customers actively buying new products) and
the imitation factor (representing customers imitating
others). Other influential factors, such as costs,
benefits, and incentives, can also be integrated into
the model.
x ABM: This approach describes customers as unique,
autonomous, and adaptive agents. These agent models
are assumed to have different characteristics, act
independently, and interact with their neighbors. The
agents' behaviors are simulated in Monte Carlo experiments,
and the model parameters are estimated
by minimizing the error between the simulated adoption
and the actual adoption data. The estimated
models are then used to perform long-term BTM
resource estimation.
To summarize these two groups of methods, ADM
describes the social dynamics of customer behaviors in a
macroscopic view. It has limitations in describing individual
customers' characteristics, but it has the advantage of
requiring fewer data. ABM is stronger than ADM in
describing individual customer's behaviors, but it requires
detailed customer data, which are often difficult to collect.
Data-Driven Control of Power Distribution
Systems With BTM Resources
As the adoption of BTM resources in power distribution
systems and buildings continues to increase, the need to
manage the operations of DERs becomes more urgent.
From the perspective of distribution system operators
(DSOs), the operation of BTM resources needs to be coordinated
to ensure system reliability. From the perspective
of building operators, the BTM resources should be controlled
to lower electricity costs and reduce electricity
service interruptions. Due to the lack of reliable distribution
network and building models, data-driven control
solutions are becoming more suitable alternatives to the
physical model-based control technologies. There has
been a tremendous amount of research and development
in the area of data-driven control for power distribution
systems with BTM resources, due to the rapid
advancement in ML. The availability of near-real-time
data from advanced metering infrastructure systems,
SCADA, microsynchrophasor measurement units, and
building control systems is making the data-driven control
technologies feasible for real-world implementation.
We cover a promising data-driven control solution for
power distribution systems with BTM resources in the
following section.
26
IEEE Electrification Magazine / DECEMBER 2022
Data-Driven VVC of Power Distribution
Systems With BTM Resources
VVC determines the operation schedule of voltage regulating
and var control devices in the power distribution
system as well as the real and reactive set points of BTM
resources, such as solar PV systems and battery storage
systems, to improve voltage quality, reduce network losses,
and lessen wear and tear on power equipment. The
rapid growth of solar PV systems and electric vehicles
makes it difficult for DSOs to keep all voltages within
appropriate limits. VVC is typically executed within a
two-timescale framework. In the slow timescale, the set
points of voltage regulators, on-load tap changers, and
capacitor banks are adjusted on an hourly basis to
improve the voltage profile and reduce equipment wear
and tear. In the fast timescale, the active and reactive
power set points of smart inverters connected to solar PV
systems and energy storage systems are changed to finetune
voltages and further reduce network losses on a
minute-to-minute basis.
The existing model-based VVC methods can be categorized
into two schemes: centralized and distributed. In the
centralized approach, the central controller, which has a
complete model of the distribution network, collects measurements
of the distribution feeder and BTM resources
and makes control decisions autonomously. Typical methods
used in the centralized VVC scheme include deterministic
optimization and robust optimization. In the
distributed approach, each voltage regulating device can
sense local grid conditions and communicate with neighboring
equipment to make coordinated VVC decisions. It
is challenging to deploy physical model-based VVC algorithms
in distribution circuits, due to the lack of accurate
distribution network models. To tackle this challenge,
data-driven methods are developed, which learn to
choose control actions based on the operational data. In
practice, many electric utilities, such as Southern California
Edison, are either considering or already switching to
data-driven VVC solutions.
Reinforcement learning (RL) is emerging as one of the
most promising data-driven approaches to solve the VVC
problem, which is essentially a sequential decision-making
task. RL teaches the controller(s) to make a good
sequence of voltage regulating decisions, which yield
proper control outcomes by using historical and/or realtime
operational data. RL leverages the framework of the
Markov decision process (MDP) to define the interactions
between a learning agent (e.g., the volt-var controller)
and its environment (e.g., the power distribution grid).
The volt-var controller and the distribution system environment
interact at each point in a sequence of discrete
time steps, as depicted in Figure 3. At each time step, the
volt-var controller collects the state of the distribution
system and, on that basis, selects the control set points
of voltage regulating devices and BTM resources. One
time step later, in part, as a consequence of the control
IEEE Electrification - December 2022

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