Integrating Machine Learning with Intelligent Control Systems for Flow Rate Forecasting in Oil Well Operations

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper studies the application of ML models to enhance the predictive capabilities of SCADA systems for oil well operations. A dataset from SCADA system is used to train and test the FlowRate and detect potential operational anomalies. It underscores the potential of integrating ML with SCADA systems to enhance operational efficiency and predictive maintenance in the oil and gas industry. 

Here are some comments.

1, the original data from SCADA system is with noise. Is the dataset pre-processed?

2, figure 5 exhibits a discrepancy with the content. 

3, the introduction part is not clear and sound.

4, the use of ML is interesting. But the dataset is too small for a real time system. What will it be with the working condition changed?

Comments on the Quality of English Language

English is good.

Author Response

Comment 1: The original data from SCADA system is with noise. Is the dataset pre-processed?

Response 1: Thank you for pointing this out. We agree with this comment and have revised the following sections of the paper to address this point: 

- Lines 142-149, 150-157, 192-196, 208-218.

Comment 2: Figure 5 exhibits a discrepancy with the content. 

Response 2: We agree with this comment; therefore, we have replaced Figure 5 with Figure 1 on page 4.

Comment 3: The introduction part is not clear and sound.

Response 3: We agree that the introduction was clear. We have entirely rewritten the introduction section for clarity and precision, with more literary works included. Lines 30-140.

Comment 4: The use of ML is interesting. But the dataset is too small for a real time system. What will it be with the working condition changed?

Response 4: We agree that this was not discussed. We have added a discussions and limitations section (4), page 13 that addresses these points.

Reviewer 2 Report

Comments and Suggestions for Authors

REVIEW

on article

Integrating Machine Learning with Intelligent Control
Systems for Flow Rate Forecasting

 

Bibars Amangeldy, Nurdaulet Tasmurzayev, Shona Shinassylov,

Aksultan Mukhanbet and Yedil Nurakhov

 

SUMMARY

The article submitted for review is devoted to an important topic. It explores the integration of machine learning with intelligent control systems for flow forecasting. This chosen topic corresponds to the focus of the Automation journal.

The study is important because the integration of machine learning with supervisory control and data acquisition systems represents a significant advance in oil well monitoring, enabling real-time maintenance prediction and operational optimization. The importance of this research is also due to the current focus on the application of advanced machine learning models, including long short-term memory, bidirectional long short-term memory, and other aspects of oil well operations. The authors used a very large dataset of 21,644 records. They obtained models that were trained and tested to predict future flow rates and detect potential anomalies.

Overall, the authors' results showed the potential of integrating machine learning with SCADA systems to improve operational efficiency and predictive maintenance in the oil and gas industry. Directions for future research are outlined and the importance of the research for modern economics and economics is emphasized.

The reviewer believes that the article touches on a very interesting and relevant topic and contains a lot of new interesting data. However, the article has a number of shortcomings that now look quite critical. These shortcomings need to be corrected. The article's shortcomings and reviewer's comments are listed below.

COMMENTS

1. The title of the article states that control systems for flow prediction were studied. Based on the title, it is not clear what industry the study was conducted for. It should be added to the title of the article that this is applicable to oil well operations.

2. The authors in their abstract report only a well-known fact about what the integration of machine learning with dispatch control and data collection systems is. At the beginning of the abstract, a narrower topic should be outlined. It must be said what a problem and scientific deficit exists in the field of this integration.

3. The abstract does not contain quantitative expressions of the results obtained. All that is said is that a new model has been proposed, that the potential for integrating machine learning with SCADA systems has been identified, and that operational efficiency and preventative maintenance in the oil and gas industry can be increased. But it is necessary to show what quantitative gains are expected from such implementations, because without this, the scientific and practical significance of this article is unclear.

4. The literature review carried out in the “Introduction” section is represented by 22 literary works. However, much more scientific and engineering works are devoted to machine learning and the study of the oil and gas industry, especially flow parameters. The authors need to work more seriously with the scientific literature. It is necessary to use more literary sources, 35-40 at least.

5. The “Materials and Methods” section is not presented in much detail. It should be provided with a flowchart with a research program. In this flow chart, it is necessary to clearly indicate what studies were conducted, what indicators were determined, and what factors were selected as variables.

6. Several images, such as Figures 4.5, 8, 9, 10 and 11, are of low quality and contain unreadable characters. This is unacceptable, they should be brought in higher quality.

7. There is a proposal to combine paragraphs 2 and 3. The “Experiments and Results” section is not sufficient to talk about contributions to science. It is also necessary to highlight a separate “Discussion” section, in which a detailed comparison of the results obtained with the results of similar studies should be provided.

8. Perhaps the authors should add more analytics to their research. It is proposed to provide an Ishikawa diagram in which to analyze all the factors affecting the effectiveness of integrating machine learning with intelligent control systems for forecasting costs in the oil industry.

9. The conclusions are presented incorrectly. It is necessary to number conclusions and provide clear results. 1 – scientific result, 2 – applied result, 3 – recommendations for real companies and 4 – prospects for the development of this research in the future.

 

10. The reference list of 24 papers is very small for research published in world-class journals. It should be increased to 35-40 items.

Comments on the Quality of English Language

Minor editing of English language required

Author Response

Comment 1: The title of the article states that control systems for flow prediction were studied. Based on the title, it is not clear what industry the study was conducted for. It should be added to the title of the article that this is applicable to oil well operations.

Response 1: We agree with this comment. The title of our article has been changed to “Integrating Machine Learning with Intelligent Control Systems for Flow Rate Forecasting in Oil Well Operations” to address this comment.

Comment 2: The authors in their abstract report only a well-known fact about what the integration of machine learning with dispatch control and data collection systems is. At the beginning of the abstract, a narrower topic should be outlined. It must be said what a problem and scientific deficit exists in the field of this integration.

Response 2: We agree with this comment. Accordingly, we have revised the abstract. Furthermore, more quantitative results have been added. Lines 11-24.

Comment 3: The abstract does not contain quantitative expressions of the results obtained. All that is said is that a new model has been proposed, that the potential for integrating machine learning with SCADA systems has been identified, and that operational efficiency and preventative maintenance in the oil and gas industry can be increased. But it is necessary to show what quantitative gains are expected from such implementations, because without this, the scientific and practical significance of this article is unclear.

Response 3: We agree. More quantitative results have been added within the abstract as found in lines 16-21.

Comment 4: The literature review carried out in the “Introduction” section is represented by 22 literary works. However, much more scientific and engineering works are devoted to machine learning and the study of the oil and gas industry, especially flow parameters. The authors need to work more seriously with the scientific literature. It is necessary to use more literary sources, 35-40 at least.

Response 4: The entire introduction/literature review section has been rewritten for clarity and the article now contains a total of 37 references.

Comment 5: The “Materials and Methods” section is not presented in much detail. It should be provided with a flowchart with a research program. In this flow chart, it is necessary to clearly indicate what studies were conducted, what indicators were determined, and what factors were selected as variables.

Response 5: We agree with this comment. We have included a flowchart for the materials and methods section, as seen in Figure 1, page 4. 

The features are casingPressure and DateTime. The target variable is Flowrate. This has been further emphasised in lines 184-189.

The materials and method section has been revised to include more details. 

Comment 6: Several images, such as Figures 4.5, 8, 9, 10 and 11, are of low quality and contain unreadable characters. This is unacceptable, they should be brought in higher quality.

Response 6: We agree with this comment. Figure 5 has been deleted and replaced with Figure 1 which depicts the flowchart of the research. Figures 8, 9, 10, and 11 quality has been enhanced. 

Figure 4 has been changed to Figure 5, and the quality of the image has been enhanced.

Comment 7: There is a proposal to combine paragraphs 2 and 3. The “Experiments and Results” section is not sufficient to talk about contributions to science. It is also necessary to highlight a separate “Discussion” section, in which a detailed comparison of the results obtained with the results of similar studies should be provided.

Response 7: We agree with this comment. Paragraphs 2 and 3 have been combined under the title “Experiments and Results”, as proposed. We have included a discussion and limitations section on page 13.

Comment 8: Perhaps the authors should add more analytics to their research. It is proposed to provide an Ishikawa diagram in which to analyze all the factors affecting the effectiveness of integrating machine learning with intelligent control systems for forecasting costs in the oil industry.

Response 8: We agree with this comment. Accordingly, an Ishikawa Diagram has been added in Figure 12 in the discussion and limitation section. Lines 447-453

Comment 9: The conclusions are presented incorrectly. It is necessary to number conclusions and provide clear results. 1 – scientific result, 2 – applied result, 3 – recommendations for real companies and 4 – prospects for the development of this research in the future.

Response 9: We agree with this comment. Accordingly, the conclusion has been revised, and a paragraph is dedicated to each point. This change can be found in lines 476-498, page 15.

Comment 10: The reference list of 24 papers is very small for research published in world-class journals. It should be increased to 35-40 items.

Response 10: We agree with this comment. The literary work has been updated to contain 37 references.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

ML method is well studied and widely used. These kind of methos are noise sensitive. I would like to know how the data are processed so that the model is effective in most conditions.

Just a simple using of the method is not an innovation.

Author Response

Comments 1: ML method is well studied and widely used. These kind of methos are noise sensitive. I would like to know how the data are processed so that the model is effective in most conditions.

Just a simple using of the method is not an innovation.

Response 1: Thank you for pointing this out. We agree with this comment and you can find a response to your comment in the article on lines 198-209.

And we also wanted to respond to you more extensively in this section:

Dear reviewer, you raise valid concerns regarding the potential impact of noise on model performance. While it is true that ML models can be sensitive to noise, training on datasets that reflect real-world conditions, including noise, is often beneficial for generalization. Removing noise entirely can lead to models that perform poorly when deployed in real-world settings, where noise is inevitable. In fact, many studies utilizing idealized, noise-free datasets artificially introduce noise (e.g., Gaussian noise) to improve the real-world applicability of their models. Our approach aligns with this principle, as we aim to develop models that are robust to the noise levels expected in real oil well data. Yue et al demonstrated that while denoising is beneficial for highly noisy data, particularly in oil reservoir settings, applying it to datasets with inherently low noise levels might inadvertently remove subtle but valuable information, potentially degrading the model's predictive power. [1]

The strong generalization we observe, evidenced by the consistently high R² scores on both the training and test sets, suggests that our models are effectively handling the inherent noise in the data. This robustness is likely enhanced by the architecture of LSTMs. The gating mechanisms within LSTMs, particularly the forget gate, allow the model to selectively discard irrelevant or noisy information over long sequences, focusing on the most salient temporal dependencies.

However, we acknowledge that excessive noise can hinder any model's ability to learn meaningful patterns. To mitigate this risk and ensure robustness even in highly noisy scenarios, it is possible to utilize noise filtering techniques such as rolling mean, moving average filters, Savitzky-Golay filters, or wavelet denoising [2]. Furthermore, implementing techniques like dropout regularization or adversarial training during the training process can further improve the model's resilience to noise.

During our experiments, we conducted preliminary sensitivity analyses by comparing model performance with and without explicit noise removal during preprocessing. Interestingly, removing the noise did not yield any significant improvements in predictive accuracy. In fact, our models achieved excellent variance explainability (R² scores consistently above 0.97) even with the noise present in the dataset. This finding further supports our approach of training on data that reflects real-world conditions, including a realistic level of noise. As a result, the noise present in the dataset was not cleaned.

[1] Yue, M.; Dai, Q.; Liao, H.; Liu, Y.; Fan, L.; Song, T. Prediction of ORFs for Optimized CO2 Flooding in Fractured Tight Oil Reservoirs via Machine Learning. Energies 2024, 17, 1303, doi:10.3390/en17061303.
[2] Iskandar, U.P.; Kurihara, M. Time-Series Forecasting of a CO2-EOR and CO2 Storage Project Using a Data-Driven Approach. Energies 2022, 15, 4768, doi:10.3390/en15134768.

Reviewer 2 Report

Comments and Suggestions for Authors

All my comments were considered.

I recommend the article for publishing. 

Author Response

Comments 1: All my comments were considered. I recommend the article for publishing.

Response 1: Dear Reviewer, thanks for your reply

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

The authors answered all my questions.

It is OK for the paper.