ETA 2021 Strategic Plan - Flipbook - Page 78
conjoined genetic engineering and bioreactor
control, such as in biofuel production. This
initiative will use an “explainable AI” strategy to
identify and then optimize the emergent system
controls. This involves fitting surrogate models
to (computationally expensive) simulations,
and then interrogating the fitted surrogates
to discover instances of (nonlinearly) coupled
controls. This approach, notably, also will
enable acceleration of policy optimization (e.g.,
through reinforcement learning) by providing
a computationally inexpensive surrogate for
physics-based simulations.
Currently, methods exist to identify control
parameters; for example, from deep neural
networks using second order methods and
forward regression. These techniques become
intractable when initial control parameters
number in the millions. Application to
integrated energy systems will serve as a
driver for methods development in explainable
AI strategies for surrogate models and
reinforcement learning.
Another challenge we will face in the use of the
testbed models is considering the importance of
extreme scenarios for energy reliability. Extrema
are, by definition, rare, and simultaneously
optimizing surrogate models to perform
under basal and exceptional conditions and
quantifying uncertainty and effect size will, even
at exascale, pose fundamental challenges in
applied mathematics. This is due in part to a
fundamental property of data-driven models:
models are unreliable far from the support
on which they are trained. An outcome of this
first year will be to develop AI models with new
inferential and uncertainty analysis procedures
for both standard and extreme conditions.
76
|
E TA S t r a t e g i c P l a n 2 0 2 1 - 2 0 3 0
Cyber-physical System Security and
Stability via Machine Learning
Cyber-physical systems, such as the U.S. electric
grid, are an interleaving of physics-based
dynamic processes with analog and digital
sensing and control systems. As a result of
this blending, the stability (in the differential
equation sense) of cyber-physical systems
is governed by the interaction of properties
from both the physical and cyber-based
components. To maintain the stability of the
overall dynamic system, proper relationships
between properties of the physical systems and
parameters of the cyber systems must be strictly
enforced. Unfortunately, these relationships are
increasingly difficult to mathematically articulate
as the number and sophistication of cyber-based
systems grow.
This initiative will develop a toolset needed
to characterize stability of high-dimensional,
nonlinear, uncertain dynamic systems as a
function of physical properties (e.g., mass) and
cyber properties (e.g., gains in control loops)
without ever having to solve the system itself.
To do so, this research will utilize techniques
from machine learning to develop algorithms to
collapse n-dimensional dynamics into a scalar
function.
The research will consist of four tasks. Task 1
focuses on the development of a high-fidelity
simulation of a microgrid with high levels
of renewable penetration. Unlike modern
simulation electric grid simulation tools, the
proposed effort will explicitly model the internal
control loops of DC/AC inverter power electronic
devices and the interaction of these devices with
the electric grid. Such a simulation will serve as
a testbed for use case generation, exploratory
analysis, and algorithm validation. Task 2
focuses on the development of Neural Lyapunov
functions using machine learning techniques to
characterize system energy and energy rate of
change as a function of system parameters. In
Task 3, research will focus on the development
of convex approximations to Neural Lyapunov
