POWER November 2021 - 25

CONNECTED PLANT
between levels), tramp air split (percent
of secondary air fed through burners not
firing biomass), and percent flue gas recycle
(FGR) back into the boiler. Based
on OPG data, the operating range for
each of these four parameters was determined
and an optimized design (42
samples) in this four-dimensional space
was performed.
The set of simulation output data
required for the digital twin grew over
time to include 60 variables grouped into
three general categories:
■ Variables Measured at the Plant. O2
outlet concentration, gas temperature
at the top of the boiler and at the outlet,
net power, and flame scanner signal
strength (15 total, one per burner).
■ Variables Needed for Constraint and
Optimization of the Digital Twin Based
on Feedback from OPG Personnel. CO
concentration at the top of the boiler
and at the outlet, NOx outlet concentration,
and maximum surface temperatures
of tube banks in the boiler and
in the convective pass.
■ Other Variables of Interest. Wall heat
flux profile (20 total), mass deposition
on walls, and total flue gas flowrate.
An instrument model for each type
of data was developed. For example,
the gas temperature at the top of the
boiler was computed by averaging
the temperatures along a line of sight
through the computational domain. The
outlet species concentrations were determined
by averaging across the outlet
plane. The flame scanner data were
extracted as the radiative intensity arriving
from the orthogonal direction at
a location just above the burner inlet. All
data were averaged over an appropriate
simulation time.
For visualization purposes, three-dimensional
images were generated that
span the range of variables and visualization
types: slice (horizontal and vertical),
surface (on all surfaces within the
boiler), and volume (volume-rendered
image of entire domain or with two cutting
planes through the domain). This
" Simulation Visualizer " is part of the
digital twin and can be accessed online
at: bit.ly/Atikokan.
Validation and Uncertainty
Quantification
The digital twin uses science-based
Bayesian machine learning (ML) methods
to combine the surrogate models
with plant measurements to update
November 2021 | POWER
the model and scenario parameters to
produce a dynamic, science-based predictive
tool that updates as the plant
evolves. For the Atikokan digital twin,
the machine learning technology receives
129 inputs from plant operations
every five minutes. It then takes the last
24 hours of data (37,000 inputs) to learn
only the continuously optimized quantities
of interest, but also the continuously
learned uncertainty in those quantities.
For the initial development project,
the digital twin was performed offline
for plant data from December 2019 and
then applied to any set of operating conditions
to produce the predictive evaluaThe
digital twin uses science-based
Bayesian machine learning (ML) methods
to combine the surrogate models with plant
measurements to update the model and
scenario parameters to produce a dynamic,
science-based predictive tool that updates
as the plant evolves.
the new state of the predictive science-
based model.
This is a Bayesian inverse problem,
not a data-fitted neural network. The digital
twin uses formal decision theory in
the presence of the learned uncertainty
to optimize the operational setpoints
for the objective function(s) subject to
the operational constraints. This decision
theory takes advantage of both a
cost function and a utility function that
accounts for the level of risk the plant
is willing to take given the uncertainty
in the system. The optimization is performed
every three to five minutes
matching the rapidly changing power
dispatch. The digital twin produces 106
outputs with uncertainty distributions
for each of the 10 level pairs for a total
of 1,060 outputs every five minutes.
These are the optimized outputs from
the twin and thus the inputs to the plant
from the twin.
A unique feature of the digital twin
is that it is continuously learning simultaneously
from experimental data as
well as from the hierarchically validated,
science-based, high-performance computing
Arches/Uintah model. Thus, the
digital twin is more than a virtual representation
that serves as a real-time
digital counterpart of a physical object or
process; it is also predictive. Our abstraction
for a digital twin is to bring together
science-based models with measurements
from the physical twin, to not
simply produce a control algorithm, but
to predict what the setpoints should be
to achieve continuous optimization of
the unit and to do so in the presence of
uncertainty. The digital twin provides not
www.powermag.com
tions. This " Offline Integration " tool can
be accessed online at: bit.ly/Atikokan.
The user may query the ML with a series
of sliders for any set of operating conditions
within the operating space of the
design of experiments described above.
Optimization with Artificial
Intelligence
One of the key lessons learned from
interactions with OPG personnel is that
uncertainty itself has little value to decision-makers
unless the tools to make
decisions-to optimize in the presence
of uncertainty-are coupled with the
uncertainty analysis itself. This is the objective
of the digital twin: to perform calculations
in real-time using the surrogate
model, and the results from the Bayesian
analysis, to make predictions with
uncertainty, and to incorporate formal
decision theory to provide optimized decisions
that account for the uncertainty.
The optimization is performed with the
use of a cost function. The cost function
translates the physics QOI (power, NOx
concentration, deposit thickness, etc.)
to a value useful to the decision-maker
(cost). This cost function is defined in
collaboration with OPG and includes the
constraints on the decision.
However, the cost function alone is
not the final function that is maximized.
As E.T. Jaynes wrote in the book Probability
Theory: The Logic of Science, " An
essential thing which is still missing ...
is the rule by which it converts its final
probability assignment into a definite
course of action. " This utility function incorporates
the risk aversion propensity
of OPG personnel. For the Atikokan digi25
http://www.bit.ly/Atikokan http://www.bit.ly/Atikokan http://www.powermag.com
POWER November 2021

Table of Contents for the Digital Edition of POWER November 2021
