Dynamic Modeling and Online Updating of Full-Power Converter Wind Turbines Based on Physics-Informed Neural Networks and Bayesian Neural Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript puts forward an interesting hybrid Physics-Informed / Bayesian Neural Network (PINN-BNN) surrogate for tackling electromagnetic-transient (EMT) studies of full-converter PMSGs. The fundamental idea—embedding nine distinct physical residuals within the loss function and then using an online mechanism to fine-tune layers flagged as uncertain—certainly has practical appeal, and the numerical results presented do look encouraging at first glance. That said, I find that several claims currently stretch beyond the evidence provided. I would be open to reconsidering a revised version, but only if the authors can convincingly address the major points outlined below.

The introduction posits that "no research has yet been conducted on electromagnetic transient simulations at the micro-second level." This statement seems to overlook some of the authors' own citations, which include Koopman-based and symbolic PINN studies from 2023–24 that indeed venture into EMT territory. It's therefore crucial for the authors to articulate, in straightforward terms, what fundamentally new capability or insight their specific approach delivers that sets it apart from these recent works.

The models are trained on a set of 500 simulated trajectories, each with a 50 µs step size. While this captures certain scenarios, real-world wind turbines experience a far wider and more unpredictable range of grid disturbances, complex control interactions, and parameter drift due to aging or environmental shifts. The current test cases, which bear a strong resemblance to the training set, don't offer sufficient proof of the model's robustness in the face of truly novel conditions. To substantiate claims of generalization, the authors should subject their model to genuinely out-of-distribution scenarios, such as deep and prolonged voltage sags, events that might trigger PLL loss-of-lock, or conditions reflecting significant aging-related parameter shifts.

The proposed online fine-tuning strategy, which adjusts layers when their output variance crosses a predefined threshold, is an intriguing concept for adaptivity. However, the manuscript provides no argument or evidence to assure us that these local weight changes successfully keep the network operating on the low-residual manifold that the PINN was painstakingly trained to respect. A brief stability analysis or a numerical check—demonstrating, for instance, that fundamental conservation laws are upheld post-update—would greatly reassure readers that these online adjustments don't inadvertently inject non-physical artifacts into the simulation.

The reported accuracy focuses predominantly on DC-link voltage and RMS stator voltage, as detailed in Tables 1–3 and Figures 2–4. For a surrogate model intended for EMT applications, a much tougher and more revealing test lies in its ability to accurately capture instantaneous stator and rotor currents, their harmonic spectra, torque ripple, and the dynamic behavior of the PLL, especially during and immediately following fault conditions. Including these more demanding metrics—and ideally, benchmarking the proposed model against an established, physics-based EMT equivalent—would make the performance claims far more compelling and provide a clearer picture of the model's true capabilities and limitations.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents a solid and original approach to the dynamic modeling of full-power converter permanent magnet synchronous generator (PMSG) wind turbines, combining Physics-Informed Neural Networks (PINNs) with an online updating mechanism based on Bayesian Neural Networks (BNNs). The study stands out for its methodological rigor, practical relevance, and innovative contribution to the field of renewable energy systems.

By incorporating physical constraints directly into the neural network loss function, the authors enhance the interpretability of the model while reducing its reliance on large datasets. The proposed online updating strategy—driven by BNNs and cosine similarity—provides a practical and effective means of adapting the model to changing operating conditions, which is particularly valuable in wind power applications.

The experimental setup is well-designed, and the results are clearly presented. The comparisons with baseline models such as ResNet and LSTM are meaningful and support the performance gains claimed in the paper.

Suggestions:

* It would be helpful to briefly explain in the introduction why CloudPSS was selected as the simulation platform, especially for readers unfamiliar with it.

* Although the PINNs-BNNs methodology is well described, including a diagram that summarizes the architecture and data flow (e.g., input variables, residuals, BNN adjustment) would further clarify the approach.

* Section 5 could benefit from a short discussion on the computational cost or training time of the proposed model in comparison to standard deep learning methods.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Dear Authors, 

I have some comments about your submitted manuscript. I hope you find them helpful. I suggest revision for you manuscript.

Let me begin by noting that the quality of the English writing is very good from my point of view, and the paper is generally clear and readable. Nonetheless, there is still room for improvement in terms of clarity and presentation.

1. I am not sure that the abbreviations used in the title (e.g., PMSG, BNN, PINN) are immediately clear to reader. I would recommend revising the title to use more descriptive terms or to minimize the use of abbreviations.
2. In equation (1), is u(t)' a typographical error, or are you referring to the derivative of u(t)?Please make it clear if it refer to derivation.
3. n equation (3), you include a summation over "n", but "n" is not defined or present in the formulation. This needs probably further clarification or modification.
4. I believe it is important to include at least one reference in the first part of Section 2, where you discuss Harmonic State Space (HSS), Linear Time-Periodic systems, and their transformation into Linear Time-Invariant form.
5. The formulation requires a clear and consistent notation section. It is currently unclear what certain expressions mean—for instance, the use of square brackets [], the meaning of diag[⋅], the distinction between bold and non-bold symbols, and the purpose of the asterisk (*) symbol. Additionally, the spaces to which the state variables belong are not specified. There is also inconsistency in notation, such as between "udc_gsc" and u_{dc_gsc} in (6) and the explanation after it, which should be clarified. Furthermore, the \Gamma function used in equations (7) and (8) is not defined. Overall, the formulation would be clearer and more rigorous with a dedicated section describing all symbols and notations. Also, in figure capital U is used. Are they different or the same?
6. I believe it would be helpful to provide more information and details about the structure and type of the artificial neural network used in your study.
7. You have stated, "However, no research has yet been conducted on electromagnetic transient simulations at the microsecond level." A deeper discussion of why this gap exists would strengthen your argument—whether it stems from the complexity, the novelty of the direction, or perhaps a previously limited recognition of its importance. Additionally, it would be valuable to explain more clearly why simulations at the microsecond level are significant, especially in the context of modern power systems and renewable energy technologies, such as wind turbines with power electronic interfaces. As a reader, I believe that the mere absence of prior research in a given direction does not, by itself, guarantee the significance of the contributions.
8. I am a bit unclear. Does this modeling approach aim to be control-oriented, or is it purely model-oriented without a control objective?
9.  In the conclusion, you have written "method not only preserves the stability of the equivalent model...". What do you mean by stability precisely? Since it can refer to different objectives in different contexts. Moreover, I have doubt that just by doing simulation, we can claim about preserving stability in general.

I hope you find my comments helpful.

Best Regards,

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Taking into account the author's revisions, answers, and explanations, I think the manuscript is suitable for publication. 

Reviewer 3 Report

Comments and Suggestions for Authors

Dear Authors,

Thank you for addressing my comments.

 Bests,