Prediction of SO_x-NO_x Emission in Coal-Fired Power Plant Using Deep Neural Network

Reviewer 1

The paper presents the results of studies on the development of a predictive model of SOx and NOx emissions when burning a mixture of various coals. In the work, the authors have demonstrated a thorough approach to the organization of the study and the selection of the analyzed parameters by the relevance of conditions. The authors had justifiedly chosen the Robust Scaler method as a method of normalizing the data, and the function of activation of Relu neuron, which allowed to solve the problem of endangered gradients (Fix The Vanishing Gradents Problem) in DNN. A comparison of the forecasting results for three ML methods by forecast (Validate the Prediction Performance) has been performed. The DNN method has given the best prediction result. There are comments on the work. 1. Low quality of drawings 1 and 2 2. The work does not give the reasons for choosing the structure of DNN (the number of layers and the number of neurons in the layer). How is the choice of DNN structure justified?

Thank you for your valuable time on reviewing the proposed paper. I appreciate the effort that you dedicated to providing feedback on our manuscript and are grateful for the insightful comments on and valuable improvements to our paper. I have incorporated most of the suggestions made by your comments. Those changes are highlighted within the manuscript in red. Please see below point-by-point the response to your comments and concerns.

(1) Low quality of drawings 1 and 2

As you commented, Figure 1,2 and 3 are modified so as to be cleared by redrawing.

(2) The work does not give the reasons for choosing the structure of DNN (the number of layers and the number of neurons in the layer). How is the choice of DNN structure justified?

Thank you for your insightful comment. We explored multitude of parameter combinations with search algorithm such as random, grid, bayesian search, not just those listed in the paper. However, no significant results were obtained through search algorithms. Among searched many configurations, we selected the parameter which give the highest performance through trial and error. Therefore, the used DNN in this paper is the most effective to address predicting emission of SOx and NOx among our trials.

(3) Third, the English writing aspect needs to be further revised.

As you commented, we have revised the whole paper and change the grammatically misunderstood parts.

Reviewer 2

Many thanks for your submitted paper. It was interesting but you need to consider many aspects for this paper and it is not as easy as you said! Please find below comments on your work:

Thank you for your valuable time on reviewing the proposed paper. I appreciate the effort that you dedicated to providing feedback on our manuscript and are grateful for the insightful comments on and valuable improvements to our paper. I have incorporated most of the suggestions made by your comments. Those changes are highlighted within the manuscript in red. Please see below point-by-point the response to your comments and concerns.

(1) First of all your paper must be improved from many aspects such as English, Structure and Scientific details

- As you commented, we have revised the whole paper and change the grammatically misunderstood parts.

(2) I can not see any novelty in your work just doing neural network!!!!

-Thank you for your kind feedback. In real industrial field, the decision regarding coal mixture ratios is often made by analyzing coal properties without systematic process. To transite towards a more systematic approach, we aim to employ an optimization model for identifying the optimal mixture ratio that minimizes SOx and NOx emissions. However, in order to determine the most suitable mixture ratio, it's crucial to comprehend the extent of SOx and NOx generation based on the properties of the mixed coal. To do this, we developed predictive model of SOx and NOx under mixture coal properties. Unlike conventional method, we believe that the prediction of Sox and NOx has novelty and it will help field engineers’ decision. In our future work, the developed model will be integrated into a reinforcement learning framework for emission control based on this model. Moreover, we are planning to continuously enhance the model's performance by incorporating additional variables.

(3) If you are talking about coal you need to provide the coal details such as proximate and ultimate analysis.

- The considered coal properties in this paper are described in Table 1. Since the coal is used as mixture, lots of different coals are used in the training process. Each coal comes from different sources, they have different properties. Moreover, in the field, the limited properties (described in Table 1) is considered so as to decide coal mixture. Therefore, our model focused these properties.

(4) When you talk about nox or sox you need to add which mechanism of them! You need to provide Engineering data

- As you commented, we add mechanism of NOx and Sox in line(42-50).

(5) When you certify your data with your field industry doesnt mean you certify with them all!!!

- As you commented, the application of our model in the field doesn't necessarily mean it's certified. As we have mention in question (2), the decision regarding coal mixture ratios is often made by human without systematic process. The decision by human experience has given no proof of usefulness. The proposed model, even it is straightforward, gives consistency in predicting Sox and NOx in the field. Therefore, we believe that it can be valuable and useful in real-world applications within the specific industry we intend to apply it to.

Reviewer 3

Thank you for your valuable time on reviewing the proposed paper. I appreciate the effort that you dedicated to providing feedback on our manuscript and are grateful for the insightful comments on and valuable improvements to our paper. I have incorporated most of the suggestions made by your comments. Those changes are highlighted within the manuscript in red. Please see below point-by-point the response to your comments and concerns.

(1) My biggest concern is from 3.2.1 data collection section, regarding the dataset quality.

- I think that your concern about the dataset quality, in the 3.2.1 data collection section, is due to the limited quantity of data and the data that couldn't be collected due to operational interruptions. We carefully eliminate the wrongly collected data and useless data. Due to data cleansing process, the available amount of data is reduced. However, with the given data, the model gives reliable results and we think the boiler does not show much deviation according to operation. The developed model is regarded as applicable in the field so that we think the model seems to be valuable and useful in real-world scenarios."

(2) The paper indicated 8 hours of unit operation data is collected, and then averaged to produce one data point for the DNN modeling. 

 It should be noted that the NOx and SO2 emissions measured at the stack are impacted by many factors, starting from the coal properties; then the boiler operation status (unit load, excess O2,  LNB (low NOx burners) settings like the primary and secondary air ratios,  tilts, swirls, etc;  APCD (Air pollution control devices) operation status,  like SCR or SNCR, ammonia injection rate, injection pattern, catalyst life; FGD or scrubber operations, like limestone quality.   While no detailed information has been provided on the coal fired plant configuration, it is very rare to assume that during any given 8 hours period of time,  the pertinent unit operation parameters that impact the NOx and SO2 will be kept constant enough such that the inherent correlation between coal properties and emissions can be maintained. Put is in another way,  in any 8 hours of time, there could be more and heavier weight features from the boiler and APCD processes for NOx and SO2 emissions than the coal properties, and these features, which are not considered by the DNN, could potentially skew the modeling results.

This should be addressed before any further recommendations.

- As you commented, it is indeed evident that a multitude of factors, including the boiler's operational parameters, significantly influence the levels of NOx and SOx emissions. In the targeted boiler, these operational parameters are automatically adjusted within specific set up boundaries predefined by the operators. Hence, the operational parameters are not exactly controlled during combustion. Therefore, we could not consider other aspect in prediction of NOx and Sox.

Currently, decisions regarding the coal mixture ratio in the industry are made primarily by analyzing the inherent properties of coal. Given the constraints present in this approach, our initial strategy was to develop a model that utilizes the available information on coal characteristics. By implementing a system based on optimized coal combinations, we intend to conduct a comparative analysis between the actual emission levels observed in operational environments and the projected emission levels from our model. This analysis will be instrumental in our ongoing efforts to progressively refine and enhance the model's performance by incorporating additional variables and considerations. The control of other parameters according to coal mixture will be our further work.

These aspects are written in line(402~409)