All reviews of published articles are made public. This includes manuscript files, peer review comments, author rebuttals and revised materials. Note: This was optional for articles submitted before 13 February 2023.
Peer reviewers are encouraged (but not required) to provide their names to the authors when submitting their peer review. If they agree to provide their name, then their personal profile page will reflect a public acknowledgment that they performed a review (even if the article is rejected). If the article is accepted, then reviewers who provided their name will be associated with the article itself.
This manuscript has undergone three rounds of revision, during which the authors have adequately addressed all reviewer comments. The reviewers have expressed satisfaction with the revisions, and I have personally verified the manuscript and the authors responses. I am satisfied with the changes made, and no further methodological revisions are required at this stage. I find the manuscript suitable for publication.
[# PeerJ Staff Note - this decision was reviewed and approved by Xiangjie Kong, a PeerJ Section Editor covering this Section #]
No additional comments.
No additional comments.
No additional comments.
The reviewer would like to thank the Authors for changes implemented in the manuscript, which in his opinion improved its quality. The points raised in the previous were adressed. I have no further comments.
**PeerJ Staff Note:** Please ensure that all review, editorial, and staff comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
The paper has been revised and resubmitted. In general, the quality of paper has been improved, and all my previous questions have been appropriately addressed.
-
-
The authors responded comprehensively to all of my previous remarks and substantially modified the manuscript content to adhere to them. In this revision, I still would like to raise a few issues, however.
- Comment 8 – In the revised version, the only difference between Fig. 2 and 4 is the “EAF Transformer” text. In my opinion, none of those figures is redundant and should be deleted. I would fully agree with the authors if the provided explanation: “Figure 4, while visually similar in some parts, includes additional mathematical and electrical parameters essential for readers aiming to reproduce or deeply understand our simulation approach. Thus, Figure 4 serves as a detailed technical reference, complementing the introductory nature of Figure 2 by incorporating critical numerical values and mathematical notations.” was actually referring to Fig. 5 instead of 4.
- Comment 16 – Although the paragraph with parameter definition was provided, I cannot find any difference between the old and new versions of the eq. (1) itself.
- Comment 17 – please define “collapse” even more precisely, i.e., what is the threshold value for voltage to be considered as a collapse, and for how long should it be occurring.
- Comment 20 – If (very rightfully) the authors consider impedance of the electrodes and cables, then in my opinion, those parameters should not be named as “arc impedance” since they do not refer directly to the “arc”. While measuring the voltage in the real system, of course, one has to take into account that the measured “arc voltage” also contains voltage drop occurring on circuit elements; however, if those are specifically quantified, they should not be named in such a way. I suggest naming them as e.g., “high-current circuit elements impedance” or directly “electrodes and cables impedance” or similarly.
-
-
**PeerJ Staff Note:** Please ensure that all review and editorial comments are addressed in a response letter and that any edits or clarifications mentioned in the letter are also inserted into the revised manuscript where appropriate.
**Language Note:** The review process has identified that the English language must be improved. PeerJ can provide language editing services - please contact us at [email protected] for pricing (be sure to provide your manuscript number and title). Alternatively, you should make your own arrangements to improve the language quality and provide details in your response letter. – PeerJ Staff
Basic Reporting
Strengths:
The article is well-structured and clearly organised, following the conventional format of abstract, introduction, methods, results, and conclusion.
Figures and tables are relevant and support the presented data well, particularly the experimental results and performance comparisons.
The language is generally clear and readable, with only minor grammatical and typographic errors that do not obscure the meaning.
Weaknesses:
Some terms and abbreviations (e.g., “CNN”, “deep features”, etc.) could benefit from clearer definitions upon first use, especially for a broader audience.
The literature review is limited and could be improved by citing and discussing more recent and relevant studies to position the proposed work more effectively within current research trends.
Recommendations for Improvement:
Add a brief explanation or reference for all technical terms when first introduced.
Enhance the literature review by including additional peer-reviewed sources, ideally from the last 2–3 years, to reflect the state-of-the-art context more accurately.
Experimental Design
Strengths:
The methodology for feature extraction, fusion, and classification is adequately described, with clear stages and justifications for each step (e.g., the use of transfer learning, CNN models, and serial-based fusion).
Evaluation metrics (accuracy, sensitivity, specificity, precision, etc.) are appropriate for the classification task.
Weaknesses:
The rationale for choosing specific models (ResNet101 and VGG19) and the fusion strategy could be better substantiated. Could you please consider exploring other fusion methods or architectures?
The dataset description (chest X-ray images) is brief. It would be helpful to include more details such as image resolution, preprocessing steps, and class distribution.
It is unclear if the experiments underwent repetition (e.g., multiple runs or cross-validation), which would have strengthened the reported performance's reliability.
Recommendations for Improvement:
Explain why we expect the chosen CNN architectures and fusion technique to outperform alternatives.
Provide more detail on dataset handling, including any data augmentation or stratified sampling used.
If not done already, conduct and report k-fold cross-validation or repeat experiments to establish statistical robustness.
Validity of the Findings
Strengths:
The reported results are promising, showing strong performance in detecting COVID-19 using X-ray images.
Comparative analysis with recent methods helps validate the effectiveness of the proposed approach.
Weaknesses:
The absence of statistical significance testing (e.g., confidence intervals or p-values) raises concerns about how robust or generalizable the reported improvements are.
Potential biases (e.g., overfitting due to small dataset, class imbalance) are not discussed, which may limit real-world applicability.
Recommendations for Improvement:
Include statistical significance testing to support the performance claims.
Discuss limitations such as data bias, generalizability, or clinical applicability. To strengthen credibility, consider external validation on an unseen dataset.
he paper addresses a highly relevant problem, and the proposed method appears technically sound and competitive.
However, to meet the standards of scientific rigour required by PeerJ Computer Science, the authors should provide a deeper discussion on model choice, experimental robustness, and limitations.
We also recommend minor language polishing before publication.
1) Literature Context:
The introduction inadequately contextualizes the novelty of TabNet/NODE for voltage collapse prediction. While prior works on voltage stability (e.g., Galiana, 1984; Kwatny, 1986) are cited, recent advances in DL-based stability assessment (e.g., transformer architectures, physics-informed neural networks) are omitted.
Improvement: Expand the literature review to include state-of-the-art DL applications in power system stability (e.g., Li et al., 2024 on transformer-based voltage stability). Explicitly contrast TabNet/NODE’s advantages over these methods.
[R] Li, Yang, et al. "Deep learning based on Transformer architecture for power system short-term voltage stability assessment with class imbalance." Renewable and Sustainable Energy Reviews 189 (2024): 113913.
2) Mathematical Rigor:
The exponential-hyperbolic arc model (Equations 1–2) lacks derivation details, and critical parameters (e.g.,B=10 V/cm) are stated without empirical validation.
Improvement: Provide a formal derivation of the arc model from first principles (e.g., Cassie-Mayr equations) and include error metrics comparing simulated vs. experimental arc characteristics.
3) Data Accessibility:
Raw simulation data and code links (GitHub/Google Drive) lack metadata descriptors (e.g., sampling rates, sensor specifications), hindering reproducibility.
Improvement: Publish a standardized data dictionary with units, acquisition protocols, and preprocessing steps. Ensure code repositories include version-controlled dependencies.
4) Temporal Resolution:
The 15–45-minute sampling window for EAF melting processes (Lines 93–94) fails to capture millisecond-scale transients critical for collapse prediction.
Improvement: Justify sampling rates using Nyquist criteria for EAF harmonics (e.g., 5–10 kHz for 50th harmonic) and include anti-aliasing filter specifications.
5) Model Benchmarking:
Comparisons with XGBoost/SVM use suboptimal hyperparameters (e.g., XGBoost n_estimators=2), artificially disadvantaging traditional methods.
Improvement: Re-benchmark with optimized hyperparameters (e.g., grid search for XGBoost) and include state-of-the-art baselines like transformer networks (Li et al., 2024).
6) Simulation Validation:
The MATLAB/Simulink model’s fidelity is asserted but not quantified (e.g., RMSE between simulated/experimental voltage profiles).
Improvement: if possible, add validation against physical EAF measurements using dynamic time warping (DTW) or spectral coherence metrics.
7) Adversarial Robustness:
The models’ sensitivity to sensor noise or adversarial attacks on I1/V1 inputs is untested, raising concerns for industrial deployment.
Improvement: Conduct Monte Carlo simulations with ±10% Gaussian noise injections and report accuracy degradation. Propose error-correction mechanisms (e.g., Kalman filtering).
8) Causal Explanation:
SHAP/LIME highlight statistical correlations (I1/V1 importance) but lack causal validation against known EAF collapse mechanisms (e.g., arc length dynamics).
Improvement: Perform ablation studies by perturbing arc length in simulations and validate if model predictions align with physical collapse thresholds.
9) Generalizability:
Training on a single-phase, balanced dataset (mean=0.5) may not generalize to imbalanced three-phase industrial systems with asymmetric harmonics.
Improvement: Test on public three-phase datasets (e.g., IEEE 39-bus) and include class-imbalance mitigation strategies.
None
In the reviewers' opinion, the manuscript needs a significant improvement – mainly in used language and style. Main sections’ naming and order is acceptable, however their contents should be adjusted in order to support better readability.
- Few sections contain rough jumps between paragraphs as if the following paragraphs belonged to different subsections, however there is no such division (e.g. topic jumps between lines 107-109 or 153-154).
- Sometimes pieces of information is unnecessarily repeated in following paragraphs, which should be reformulated. E.g. paragraphs between lines 276 and 294. Additionally, this some parts are repeated again around line 435.
- Electrical engineering technical texts rather do not use term “voltage crashes”, it is rather a dip, sag or collapse.
- Line 334 – “distance between the melted metal and the scrap” should rather be distance between electrodes and the scrap or melted metal and electrodes.
- Line 341 – “A value is a constant defined by the anode and cathode voltage collapse” in this context it should be not voltage collapse but rather a voltage drop.
I would strongly suggest checking the text with native English speaker with technical background because the manuscript language and readability seems to be not improved but rather hindered by the generative AI. Additionally, there are few expressions which suggest lack of knowledge about EAF operation nuances. I do not suggest that the authors lack it, but maybe application of AI-based language correction missed them resulting in a text which uses synonymous but less specific expressions.
Figure 2 is not referenced in the text. Additionally there is a typo – “Harmonik”. The schematic looks a bit rough and unprofessional. In my opinion it should be enhanced.
Figure 3 is far too general to be applied in such technical context.
Figure 4 contains duplicated part of Figure 2.
Figure 5 has lines with different width, which again seems unprofessional, it should be improved.
Figure 7 contains a screenshot from Simulink window with highlighted “m” and “Vat” blocks which is related to the fact that those variables were not defined in Matlab workspace. Consequenly presented model is not usable. Such figure should not contain any indication of errors in used software.
Figures 8-9 are not referenced in the text.
In the reviewers’ opinion the idea of voltage collapses forecasting and predicting by means of deep-learning methods is relevant and meaningful. However, the manuscript nowhere explains how applied methods are or should be actually applied for such a task. To my understanding the content of the manuscript only describes development of methods classifying previously recorded data and not predicting whether such event will occur. This is backed up by the authors themselves in line 693 – “models demonstrated robust classification capabilities”.
Additionally, there are more remarks which indicate that the research technical standard should be improved:
- Features I1 and V1 are not defined. I assume that these are electrical quantities of the “first” phase in a three phase system, but it is not explicitly defined.
- Line 96 contains information that bus resistance should be high to ensure stability which I believe is not entirely correct.
- In line 259 – “sampling and data size reduction have been used to improve prediction accuracy in voltage stability assessment” – this sentence is just not true. It is not just sampling or data size reduction that improves stability assessment. The cited research was misinterpreted.
- Equation (1) conditions are wrongly defined as they cover only rising edge of the positive half-period and falling edge of the negative half-period. At the same time the falling edge of positive half-period and rising edge of negative half-period is not covered. Some of the equations’ symbols are not defined.
- What is considered as the most important feature “collapse” (line 386) is not explicitly defined.
- Table 2 contains data which is meaningless in case of periodic signals – e.g. average value of current or voltage.
- Equation in line 436 lacks numbering.
- Table 1 contains “arc reactance” which contain arc resistance and reactance value so it should rather be arc impedance. Additionally, how is this constant value relevant to the study as arc impedance changes during the EAF operation stochastically?
The data and methods have been described but in my opinion there are not adequate to the goal suggested in the introduction. This remark relates with all of the technical issues raised in previous point. Conclusions correctly sum up what was presented in the main body of the manuscript.
In the reviewers’ opinion the manuscript contains many language, style and technical errors. Their overall weight is so significant that I must suggest rejecting the paper. I suggest rethinking the idea behind presented research and preparing completely new, redesigned version of the paper for a separate submission.
All text and materials provided via this peer-review history page are made available under a Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.