POWER November 2021 - 27

CONNECTED PLANT
Uintah) has been developed from a
formal hierarchy of targeted validation
experiments ranging from bench-scale
furnaces through pilot-scale pulverizedfuel
rigs to full-scale (1.2 GW) utility boilers.
The
advantage of
this
5. Atikokan boiler efficiency comparison for the month of December 2019 between as-operated
conditions with the digital twin-optimized operation at the same burner level and the total
digital twin-optimized operation. Averaging these efficiency data over the entire month shows
that a 1.5% improvement was noted using existing burners in service and a 6.1% improvement
was observed with optimal burners and operating conditions. Courtesy: OPG
digital
twin
model is that it can be used to predict
highly resolved, validated quantities of
interest from first principles that cannot
be measured. This model is then used
to create surrogate models to span the
range of the operational space for the
Atikokan station. These surrogate models
take seconds to run in the online
digital twin. The Atikokan digital twin has
been demonstrated in a pseudo-online
mode using dynamic plant data collected
from the month of December 2019. In
the end, the Atikokan digital twin demonstrated
that:
6. Atikokan boiler efficiency and NOx emissions comparison between as-operated conditions
with the digital twin-optimized operation at the same burner level and the totally optimized
operation. Courtesy: OPG
Weighting factors are applied to each
objective function depending upon priorities.
In this case, the power level is given
a 40% weighting factor (as there is some
small variability acceptable) and the boiler
efficiency is given a 60% weighting factor.
Thus, Figure 5 compares the boiler efficiency
under the as-operated conditions
for the month of December 2019 with
both the boiler efficiency that could have
been achieved firing with the same burners
but with optimal operating conditions,
and the efficiency that could have been
realized had the boiler been operated
with the optimum burners in service and
optimum operating conditions. Averaging
these efficiency data over the entire
month of operation shows that a 1.5%
improvement was noted using existing
burners in service and a 6.1% improvement
was observed with optimal burners
and operating conditions.
Suppose that low NOx
emissions
were equally important as boiler efficiency.
The weighting factors now become
40% boiler load, 30% boiler efficiency,
and 30% NOx emissions. The Atikokan
digital twin reoptimizes operating conditions
and burner levels to meet these
new objective function constraints.
The results are shown in Figure 6.
Notice now that there is only a 3.4% average
increase in boiler efficiency when
firing at optimum burner levels and firNovember
2021 | POWER
ing conditions compared to as-operated
conditions. However, with that 3.4%
average increase in efficiency comes a
47.0% average decrease in NOx emissions
compared to as-operated conditions.
Thus, the number and choice of
ultimate objective functions are crucial
to optimizing the operation of the boiler.
A Successful Demonstration
OPG's Atikokan power station is a 227MW,
opposed-wall-fired, wood pellet
biomass plant that is required to run at
very low loads (about 50 MW). Under
these conditions only two burner rows
are in service. Operational options include
the amount of air in the out-ofservice
burners, amount of recirculated
flue gas, and amount of excess air. An
online digital twin has been developed
for the purpose of optimizing operating
conditions to minimize operating costs
for electric power generation under dynamic
low-load (40 MW to 100 MW) conditions
at Atikokan.
This objective has been achieved
through a Bayesian machine-learning
(ML) and decision (AI) model that is continuously
updated by combining offline
data from high-performance computing,
physics-based large-eddy simulations,
and online data from the power plant's
distributed control system. The physicsbased,
biomass boiler model (Arches/
www.powermag.com
■ Bayesian physics-based AI introduces
a method to improve operator decision-making
by quantifying bias and
from first principles that would be inaccessible
to AI methods which use
experimental measurements alone.
■ Computational expenses associated
with offline large-scale simulations are
mitigated by the ability of the physicsbased
machine learning to deliver
quantities of interest that cannot be
measured directly.
■ Continuous optimal operations are
achievable by merging physics-based
models with Bayesian decision theory.
Using the Atikokan digital twin in a
pseudo-online mode with plant data
from December 2019, comparisons
were made showing operating conditions
resulting from the optimization of
various objective functions. When optimizing
only on boiler efficiency, an average
efficiency increase of 6.1% was
noted compared to as-operated firing.
However, when optimizing equally on
boiler efficiency and NOx emissions, the
average efficiency increase dropped to
3.4%, but the average NOx
emissions
decrease was 47.0% compared to asoperated
firing. The Atikokan digital twin
is anticipated to be fully installed and operational
by the end of 2021. ■
-Philip Smith (philip.smith@utah.edu)
is a chemical engineering professor at
The University of Utah, Les Marshall (les.
marshall@opg.com) is a senior technical
officer with Ontario Power Generation,
and Stan Harding (n.stanley.harding@
gmail.com) is principal of N.S.
Harding and Associates.
27

http://www.powermag.com

POWER November 2021

Table of Contents for the Digital Edition of POWER November 2021