A Hybrid Deep Learning and Model Predictive Control Framework for Wind Farm Frequency Regulation
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe title is clear but could be made more concise by moving the sustainability phrase into the abstract. The abstract is informative yet too long; shortening it and adding precise numerical results would strengthen its impact. The introduction presents good background but is dense shorter paragraphs and a clearer statement of the research gap would improve it. The methodology is well explained, though figure captions should be more descriptive, acronyms redefined when first used, and equations given brief intuitive explanations. Results are convincing, but captions could include quantitative comparisons and more baseline methods should be discussed; adding computational cost analysis would help. The conclusion should be more concise, avoid repetition, and include reflections on real-world application. Language can be improved by shortening long sentences, keeping tense consistent, and replacing vague terms with exact numbers.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper proposes a two-layer control framework combining CAE-DNN and nonlinear MPC for wind farm frequency regulation. The topic holds significant practical importance and application value. The article is well-structured and logically organized. However, there is room for improvement in the elaboration of technical innovation and some critical details.
Major Comments:
1. The authors claim to propose a "novel" hybrid framework. However, the use of deep learning (particularly DNN and CAE) for feature extraction and mapping, as well as MPC for distributed system optimization allocation, are not entirely new in the field of control. The true innovation of this paper appears to lie in the integration of these specific techniques within this specific two-layer structure to address the specific problem of wind farm frequency regulation. It is recommended that the authors more modestly and precisely define their contributions in the introduction and abstract, emphasizing the effective integration scheme rather than solely highlighting the "novelty" of individual components. This approach will prevent reviewers or readers from developing overly high expectations only to find that it is a combination of existing technologies.
2. DNNs are often regarded as "black-box" models. The paper outputs an "optimal coefficient" to determine the total power deficit. Could the authors further explain the physical or control meaning behind this coefficient? Does it relate to known physical quantities such as the system's inertia constant or damping coefficient? Enhancing the interpretability of the model would significantly deepen the work and improve its persuasiveness.
3. The performance of the DNN model trained on PSO-optimized data heavily depends on the coverage of the training dataset. Has the paper tested its performance under extreme or uncommon conditions not covered in the training set (e.g., drastic wind speed changes or compound faults)? Please supplement the discussion on the generalization boundaries of this data-driven method and potential improvement strategies (e.g., online learning, transfer learning).
Minor Comments
1. Please explicitly specify the solver used in the real-time optimization （Distributed optimization of nonlinear multi-agent systems: A small-gain approach， IEEE TAC;Global Asymptotic Stability Analysis for Autonomous Optimization, TAC; Gradient-Free Cooperative Source-Seeking of Quadrotor Under Disturbances and Communication Constraints, TIE） of the nonlinear MPC and its computational time consumption. This is crucial for evaluating the feasibility of implementing this strategy in actual wind farm control systems.
2. The position of the title in Figure 11-(e) is incorrect.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authorscomment in app
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe manuscript proposes a hybrid framework that integrates deep learning and model predictive control to regulate the wind farm frequency. The framework employs a bi-level strategy to address critical challenges from the intricate relationships that degrade dynamic performance. The upper layer adopts a synthetic inertial intelligent control strategy based on contractive autoencoder and deep neural network. The lower layer employs a nonlinear MPC strategy to allocate the total power deficit to each turbine by using the discretized rotor motion equation as the prediction model and optimizing under constraints to ensure stable and efficient frequency regulation. However, this reviewer considers that the article cannot be published due to the following comments.
Major issues
-The work has serious theoretical and technical shortcomings. In theoretical terms, it simplifies the problem of a wind turbine farm to a single wind turbine. In technical terms, it uses expressions to estimate response parameters instead of computer programs, that are more robust.
-Abstract. Add the main hypotheses to achieve the objective of the work. Also, add the main results in quantitative terms.
-Introduction. Enhance this section by adding more references and discussions on the topic.
-Lines 216-225. There are programs that can be used to model wind turbine farms. In this regard, what is the reason to assume simplifications that lead to errors or biases in the results?
-The architecture of the DNN used and the reason for its selection are not mentioned. In addition, the inputs and expected outputs are not mentioned.
-Section 5.1 What angle was considered for the blades and justify.
-Figure 10. It seems that expressions are used to estimate parameters instead of a computer program that rationally estimates the wind response of the farm. This issue is important because if the response of the farm is not estimated rationally, the input data of the proposed framework will be biased and, therefore, the results.
- Conclusions. Add conclusions on the findings in quantitative terms.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 4 Report
Comments and Suggestions for AuthorsThe manuscript proposes a hybrid framework that integrates deep learning and model predictive control to regulate the wind farm frequency. The framework employs a bi-level strategy to address critical challenges from the intricate relationships that degrade dynamic performance. The upper layer adopts a synthetic inertial intelligent control strategy based on contractive autoencoder and deep neural network. The lower layer employs a nonlinear MPC strategy to allocate the total power deficit to each turbine by using the discretized rotor motion equation as the prediction model and optimizing under constraints to ensure stable and efficient frequency regulation. However, this reviewer considers that the article cannot be published due to the following comments.
Major issues
- As this reviewer previously commented, the work presents both theoretical and technical shortcomings. These shortcomings must be solved before using artificial intelligence techniques. We must be conscious that if we simplify processes that inherently exhibit bias and then, we use artificial intelligence techniques, the result is something with greater bias.
Author Response
Comment 1: As this reviewer previously commented, the work presents both theoretical and technical shortcomings. These shortcomings must be solved before using artificial intelligence techniques. We must be conscious that if we simplify processes that inherently exhibit bias and then, we use artificial intelligence techniques, the result is something with greater bias.
Author response: We sincerely thank the reviewer for raising these critical theoretical and technical concerns, which have helped us significantly improve the rigor and clarity of the manuscript. We apologize for any lack of clarity in the original version that may have led to misunderstanding.
In response to the theoretical concern regarding simplification of the wind farm to a single turbine, we wish to emphasize that our work does not employ such a simplification. Instead, we explicitly model multi-turbine dynamics and wake interactions—a point that has now been further clarified in the revised Section 2. Specifically, the Jensen wake model with square-sum superposition is used to quantify the impact of upstream turbines on downstream units, and the lower-layer model predictive control (MPC) strategy optimizes power allocation for each turbine individually under tailored operational constraints. This ensures that the dynamic couplings among turbines are fully considered.
Regarding the technical observation on the use of expressions rather than computer programs, we clarify that the system-frequency-response (SFR) model adopted is a well-established approach in power system stability studies. It is enhanced with an embedded high-fidelity wind farm dynamic model—solved via MATLAB/Simulink and MPC—that explicitly incorporates wake effects and propagation delays, ensuring both physical accuracy and computational tractability. Relevant explanations have been added in Section 5 (Case Study).
We fully acknowledge your concern about potential bias amplification, and our revisions—by explicitly modeling multi-turbine interactions (via the Jensen wake model with square-sum superposition and turbine-specific constraints) and embedding a physically consistent wind farm dynamic model (solved via MATLAB/Simulink and MPC) within the SFR framework—have eliminated inherent biases from oversimplification, ensuring that the application of AI techniques does not lead to greater bias in the results. We hope these revisions have adequately addressed the reviewer’s concerns and thank them again for their valuable guidance.
