(749c) Dynamic Real-Time Optimization of a Coal-Fired Power Plant Using an Artificial Neural Network Model

Dynamic Real-Time Optimization of a Coal-Fired Power Plant
Using an Artificial Neural Network Model

Load demand on fossil fuel power plants is more variable
today than ever before. A major contributor to these variable load demands is
the penetration of renewable energy sources into the grid. As intermittent
renewable energy like wind and solar becomes available, the load demand to
baseload units like coal power plants must decrease rapidly to allow room for
renewable energy to enter the grid. Similarly, when these same renewables
become unavailable, base load units must ramp up quickly to make up the
difference in supply and demand^[2]. The process of continual ramping
is highly inefficient, costing between $150 million and $450 million in
wear-and-tear costs annually, and in effect increases emissions from power
generators as a whole^[6]. Due to the process’s overall complexity,
techniques used in other industries to improve efficiency (such as
first-principles modeling) are difficult to apply due to massive computational
time and limited engineering resources^[3,5]. A proven method of
reducing emissions of coal-fired boilers is the use of neural networks to model
the combustion process, and subsequent optimization of the input variables^[4].
Due to the highly nonlinear nature of the process, swarm based optimization approaches
have grown increasingly popular for combustion optimization^[1] in
place of more traditional gradient based optimizers or collocation methods.
Traditional real-time optimization (RTO) methods worked well when base load
power plants operated as they were originally designed, consistently at maximum
or near-maximum loads. With today’s variable load demands however, power plants
operate less often in steady-state configurations, or more often in ramping
conditions, attempting to meet these ever-changing demands, and as a result, traditional
RTO methods are less-effective. In many systems, by the time the optimizer is
able to operate and determine input conditions, the system has already changed
configuration to where the current solution from the optimizer is no longer
valid. The major cause of this is the increasingly dynamic nature of the
process, while still attempting to apply methods that are meant for
steady-state operation. This work focuses on improving methods of combustion
optimization by applying dynamic real-time optimization techniques (DRTO) using
a neural network model. To demonstrate the effectiveness of this configuration,
a basic simulation model representing a coal-fired power generation unit is
constructed. Treating the simulation as the power plant, a nonlinear
autoregressive exogenous (NARX) neural network model is developed using data
from the simulation as the training set. Upon development of a suitable model, constrained
model predictive control (MPC) is implemented on the unit. Emissions values and
unit efficiency from this application are compared to those produced from a
traditional RTO configuration applied to the same process. Improvements are
shown to both emissions levels and unit efficiency from the DRTO configuration
over the RTO configuration.

^[1] Song, J., et al., “Improved artificial
bee colony-based optimization of boiler combustion considering NOx emissions,
heat rate and fly ash recycling for on-line applications,” Fuel, 172,
pp. 20–28, (2016)

^[2] Gowrisankaran, G., et al., “Intermittency
and Value of Renewable Energy,” Journal of Political Economy, 124,
(4), (2016)

^[3] Thornock, D. E., et al., “Evaluation of
Carbon Conversion/NOx Emissions from Full Scale Boilers Using CFD-Based
Combustion Simulations,” presented at meeting of the American Flame Research
Committee, Maui, HI (Oct 16-20, 1994)

^[4] Chu, J. et al., “Constrained Optimization
of Combustion in a Simulated Coal-Fired Boiler Using Artificial Neural Network
Model and Information Analysis,” Fuel, 82, (6), pp. 693-703,
(2003)

^[5] Darby, M. L., et al., “RTO: An Overview
and Assessment of Current Practice,” Journal of Process Control, 21,
(6), pp 874–884, (2011)

^[6] Moarefdoost, M. M., et al., “A Robust
Model for the Ramp-Constrained Economic Dispatch Problem with Uncertain
Renewable Energy,” Energy Economics, 56, pp. 310–325, (2016)