Prediction of CO–NOx Emissions from a Natural Gas Power Plant Using Proper Machine Learning Models

Abstract

Four machine learning (ML) models including a deep neural network, a long short‐term memory network, a random forest (RF), and an extreme gradient boosting are implemented to predict CO–NOx emissions from a natural gas power plant. A new feature optimization scheme (FOS) via a sequencing process of feature selection and hyperparameter optimization can intensify the ML models. Through the procedures of training, validation, and testing, reliable ML models need to take high prediction accuracy and fast training into account. After a few comparisons, it is found that 1) the FOS effectively improves the prediction accuracy by 18%–67%; 2) the FOS‐based RF model is an appropriate option to carry out the fast and accurate prediction of CO–NOx emissions by using the decision tree classifiers.

Full text

Prediction of CO–NO
x
Emissions from a Natural Gas
Power Plant Using Proper Machine Learning Models
Wei Wu,* Yan-Ting Lin, Po-Hsuan Liao, Muhammad Aziz, and Po-Chih Kuo*
1. Introduction
The increase in CO
2
emissions has caused irreversible damage to
the earth’s ecological environment,
[1]
and CO
2
emissions from
fossil fuel and industrial processes were up to 60% of global
greenhouse gas emissions.
[2]
Recently, machine learning (ML)
technologies have been widely implemented in energy systems
in terms of power efficiency and gas emissions. A comprehen-
sive review of ML applications showed that ML technology was
an effective way to reduce power losses and increase the system
efficiency of the integrated power system using a combination of
smart grid and renewable energy sectors.
[3]
As internal combus-
tion engines played an essential role in power generation, ML
techniques using unsupervised learning, supervised learning,
and reinforcement learning could provide useful solutions for
modeling internal combustion engines.
[4]
The flue gas from coal-fired power plants contains harmful
pollutants such as SO
x
and NO
x
. Regarding the prediction of
SO
x
–NO
x
emissions from a class of power generation systems,
the support vector machine (SVM) model
was more accurate for predicting NO
x
emission than the feedforward neural
network (FNN) model,
[5]
a deep neural net-
work (DNN) via the data preprocessing and
specific feature selection could effectively
reduce the computational time for predict-
ing SO
x
–NO
x
emissions,
[6]
an ensemble
DNN model has better prediction perfor-
mance for predicting NO
x
emission,
[7]
and an adaptive network-based fuzzy infer-
ence system as a kind of artificial neural
network was validated to accurately predict
NO
x
emissions.
[8]
Besides, a neural net-
work time-series nonlinear autoregressive
and a Gaussian process regression (GPR)
have performed better in forecasting CO
2
emissions of power plants in some specific countries.
[9]
On
the other hand, a case study showed the enhanced long
short-term memory network (LSTMN) could most accurate
and stable prediction of NO
x
emission rates from a coal-fired
power plant during transient operation.
[10]
An FNN framework
integrated with a kinetic-based process was utilized to generate
life cycle inventory data from different types of woody biomass
with hundreds of characterization data samples such that the
large variations in energy consumption and greenhouse gas
emissions across different biomass species were specified.
[11]
A ML algorithm using an autoregressive moving average model
with exogenous inputs was developed to forecast the CO
2
emission intensities in European electrical power grids, where
short-term forecasts could help electricity consumers schedule
their load to minimize CO
2
emissions.
[12]
In addition, a deep
learning-based FNN model was found to be suitable for predict-
ing the amount of CO
2
emission from the specific power sector
in Kuwait,
[13]
and the SVM and DNN models were effectively
implemented to forecast transportation-based-CO
2
emission
and energy demand in Turkey,
[14]
an extreme learning machine
for predicting carbon emission intensity of some cities,
[15]
and an
FNN-based optimization using LSTMN could outperform the
steady-state optimization and improve the thermal efficiency
of coal-fired boilers.
[16]
By the above statements, the FNN using a deep learning algo-
rithm could effectively predict NO
x
and CO
2
emissions, but the
prediction performance depends on raw data features. It means
that the preprocessing step for transforming raw data into fea-
tures is a core approach to improving the performance of ML
models. In this article, the appropriate and ecologically valid data
for harmful pollutants including NO
x
and CO from a natural gas
power plant are denoted as the raw data.
[17]
The descriptions of a
natural gas power plant and four ML models are addressed in
W. Wu, Y.-T. Lin, P.-H. Liao
Department of Chemical Engineering
National Cheng Kung University
Tainan 70101, Taiwan
E-mail: weiwu@gs.ncku.edu.tw
M. Aziz, P.-C. Kuo
Institute of Industrial Science
The University of Tokyo
Tokyo 153-8505, Japan
E-mail: pckuo@iis.u-tokyo.ac.jp
The ORCID identification number(s) for the author(s) of this article
can be found under https://doi.org/10.1002/ente.202300041.
DOI: 10.1002/ente.202300041
Four machine learning (ML) models including a deep neural network, a long
short-term memory network, a random forest (RF), and an extreme gradient
boosting are implemented to predict CO–NO
x
emissions from a natural gas
power plant. A new feature optimization scheme (FOS) via a sequencing process
of feature selection and hyperparameter optimization can intensify the ML
models. Through the procedures of training, validation, and testing, reliable ML
models need to take high prediction accuracy and fast training into account. After
a few comparisons, it is found that 1) the FOS effectively improves the prediction
accuracy by 18%–67%; 2) the FOS-based RF model is an appropriate option to
carry out the fast and accurate prediction of CO–NO
x
emissions by using the
decision tree classifiers.
RESEARCH ARTICLE
www.entechnol.de
Energy Technol. 2023,11, 2300041 2300041 (1 of 12) © 2023 Wiley-VCH GmbH