Next Article in Journal
Particulate Matter 2.5 (PM2.5): Persistence and Trends in the Air Quality of Five India Cities
Previous Article in Journal
Discovery of Large Methane Emissions Using a Complementary Method Based on Multispectral and Hyperspectral Data
 
 
Article
Peer-Review Record

A Deep Learning Model for NOx Emissions Prediction of a 660 MW Coal-Fired Boiler Considering Multiscale Dynamic Characteristics

Atmosphere 2025, 16(5), 533; https://doi.org/10.3390/atmos16050533
by Jianrong Huang 1, Yanlong Ji 2,* and Haiquan Yu 3
Reviewer 1:
Reviewer 2:
Atmosphere 2025, 16(5), 533; https://doi.org/10.3390/atmos16050533
Submission received: 25 March 2025 / Revised: 15 April 2025 / Accepted: 28 April 2025 / Published: 30 April 2025
(This article belongs to the Section Air Quality)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. In the abstract, you stated that you would introduce MSGNet, but in the introduction (lines 67 to 71), you indicated that this is only an application of the method. Please ensure consistency between the abstract and the introduction.
2. What does "FFT" refer to in the abstract? Please note that any abbreviation should be defined when first mentioned, even in the abstract.
3. Some quantitative results should be provided in the abstract to better demonstrate the effectiveness of the proposed method.
4. In the literature review, there is a vast number of studies related to optimizing the ML models for improving the NOx and CO emissions predictions, but only a limited number are cited in this study. This is insufficient to establish a strong foundation. For example, Ref. [1] proposed a stacked ensemble machine learning (SEM) model for predicting CO and NOx emissions from a CCPP gas turbine. By combining multiple base learners and optimizing the model using Bayesian optimization, the study demonstrated improved predictive accuracy compared to traditional machine learning approaches reported in the literature. Ref. [2] proposed a novel hyperparameter optimization framework that combines Metropolis-Hastings and random sampling strategies to enhance the tuning of deep learning models for regression tasks. The study demonstrated that the predictive performance of the DNN model after optimization was significantly improved compared to conventional methods, such as Bayesian optimization, random search, and Hyp-RL, particularly on complex, high-dimensional datasets like NOx and CO emission. Ref. [3] investigated the performance of a novel Random Forest (RF) regression model for predicting key output variables from an experimental dataset of a diesel engine adapted to operate on both compressed natural gas and diesel fuels. By employing feature engineering techniques, hyperparameter tuning using a tree-structured Parzen estimator, and model interpretation through SHAP analysis, the study demonstrated that the RF model achieved high predictive accuracy, confirming its effectiveness and competitive performance for engine emissions prediction. These are just a few references that I have found. Authors are encouraged to explore additional related works.
5. In Section 2.1, please clarify how the arrangement of combustion system features can help reduce NOx emissions. Is this improved design proposed in this study? If not, appropriate citations are needed.
6. In Section 2.2, the authors refer to "previous research" but do not cite any papers. The reasoning behind the parameter selection should be clearly explained and supported by evidence.
7. What would happen if the data were not normalized? Please discuss this potential issue.
8. Figure 5 shows a promising reduction in parameter selection; it could reduce the computation cost and make the ML model more efficient. However, could you provide the results of using all parameters and compare them with the results after reducing to 18 input variables?

References 
[1] Pachauri Nikhil. "An emission predictive system for CO and NOx from gas turbine based on ensemble machine learning approach." Fuel 366 (2024): 131421.
[2] Tiep Nguyen Huu, Hae-Yong Jeong, Kyung-Doo Kim, Nguyen Xuan Mung, Nhu-Ngoc Dao, Hoai-Nam Tran, Van-Khanh Hoang, Nguyen Ngoc Anh, and Mai The Vu. "A New Hyperparameter Tuning Framework for Regression Tasks in Deep Neural Network: Combined-Sampling Algorithm to Search the Optimized Hyperparameters." Mathematics 12, no. 24 (2024): 3892.
[3] de Lima Nogueira Silvio Cesar, Stephan Hennings Och, Luis Mauro Moura, Eric Domingues, Leandro dos Santos Coelho, and Viviana Cocco Mariani. "Prediction of the NOx and CO2 emissions from an experimental dual fuel engine using optimized random forest combined with feature engineering." Energy 280 (2023): 128066.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Some physical and chemical properties of coal were not considered in this paper, Among them, the most important ones are nitrogen content and the particle size.

The NOx formed in boilers can be attributed to three sources, thermal NOx, fuel NOx, and prompt NOx. The nitrogen content of coal also plays an important role to determine the NOx production. However, it was not considered in this paper.

Coal was grinded to fine particles before being injected into boiler. The particle size would affect the burning speed, which would in turn affect the combustion temperature and the associated NOx. The particle size of coal was not considered in this paper either.

Since coal properties were not considered as variables of the model in this paper, the specific properties of the coal used in this power plant while the measured NOx data were collected should be mentioned clearly so that the readers may know the applicability of this model. This model is valid only when the coal properties are identical to the coal properties used in this plant. This is the limit of this model.

The meaning of the horizontal axis in Figure 2 should be explained. What does it mean for the digital number from 1 to 13771?

Usually, the training data, the validating data, and the testing data were selected randomly. However, Figure 2 shows that these data were separated in distinct regions. It is noted that the load variations in testing data are much less than that of training data. We would never know the performance of this model while load varied in high levels.

Please show the original measured data of some other control variables, such as Coal Rate A and Secondary Air B.

The oxygen content shown in Figure 2 are at very low level for a boiler combustion. Usually the oxygen content in a boiler, which is operated at lean burn combustion, should be around 10%.

A rough comparison between the trends of variation of oxygen and NO in Figure 2 show that they vary in the same trends. That is, as oxygen increases, NO would increase too. However, low oxygen means less excess air, that would result in high combustion temperature, and the associated NO would be high. 

The meaning of the horizontal axis in Figure 7 should be explained. What does it mean for the digital number from 1 to 6?

It was found that oxygen content in the flue gas plays the most important role to determine the NO emission in the boiler, implying that thermal NOx is the major mechanism of NO formation. However, oxygen content is not a controllable variable, it is the final product of coal combustion, the same as NO. If this model is going to be useful for the emission management, the variables should be controllable, so that we may control the NO emission by adjusting the setting of each variable.

Please specify the number of points in Figure 6 for all models. It seems that the number of points in SMGNet model is less than that of other models. (The straight lines look fat for other models while the straight line of MSGNet looks thin.)

This paper emphasizes a new method of modelling. The theory of MSGNet covers a lot of pages in this paper, from equ.(1) to equ.(17). It takes quite a long while to read it over. However, the advantage of this method has net been discussed thoroughly. Only the final results were compared with other models in Figure 6 and Figure 7. I would expect readers of this paper may learn much more from the discussions of this paper by the authors.

The authors confirm MSGNet’s practical applicability to enhance denitrification optimization (direct quote from the conclusion of this paper ), please convince me with one example. Please use this model to demonstrate how to control the NO emission at a stable value by adjusting the control variables while the load of boiler is changing.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for your response. I think your paper is much better and can be accepted now.

Back to TopTop