Probabilistic Wind Speed Forecasting Under at Site and Regional Frameworks: A Comparative Evaluation of BART, GPR, and QRF

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

Several areas could be refined to strengthen the paper’s clarity, reproducibility, and scientific impact. My specific comments are provided below to assist the authors in enhancing the quality and readability of the manuscript.

1. The abstract is dense with numerical detail, which may overwhelm general readers; consider simplifying or summarizing key statistics (e.g., focus on relative ranking rather than multiple decimal values).

2. The abstract could end with a concise statement of broader impact how these findings can be generalized beyond the studied region

3. The introduction is technically robust and contextually rich. However, it would benefit from tightening the narrative to emphasize the research gap, methodological contribution, and novelty more explicitly, while trimming descriptive meteorological background

4. In the Data Description and Quality Control section, it is unclear whether the interpolation was applied prior to or after standardization, which can affect normalization scales and model learning dynamics.

5. The text notes exclusion of values exceeding ±3 standard deviations, but this threshold may not adequately reflect meteorologically valid extremes, particularly for wind speed. The authors should justify whether this statistical criterion aligns with physical thresholds or BoM quality standards.

6. In the methodology section, the paper should specify how the number of trees (T) was determined, as this directly affects variance reduction and computational cost.

7. Discussion on potential bias for extreme quantiles (τ close to 0 or 1) would be valuable, as QRF tends to produce conservative tails when training data are sparse.

8. In the results section, strengthen interpretation by discussing operational implications: what does “+0.006 m/s RMSE” mean for energy yield uncertainty or forecasting reliability?

9. Add a compact statement comparing “variance capture vs. phase capture”: e.g., “QRF maintains phase fidelity (r ≈ 0.53) even when R² stagnates, suggesting robust timing accuracy even if amplitude matching deteriorates.

10. Clarify what magnitude of CRPS difference is practically significant, e.g., does a 0.1 m/s improvement translate into meaningful predictive skill in energy modeling?

11. Consider adding a summary visualization (e.g., bar chart or heatmap) showing variable importance across all models and sites.

12. In the conclusion section, reducing numeric redundancy (e.g., RMSE deltas can move to Results), Strengthening interpretive synthesis (why do these findings matter for real-world forecasting, model design, and grid integration?) and structuring into three parts: (1) Summary of findings, (2) Interpretation & implications, (3) Future research directions.

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents a comprehensive and rigorously evaluated comparison of Quantile Regression Forests, Bayesian Additive Regression Trees, and Gaussian Process Regression for probabilistic wind speed forecasting. The study is well-designed, leveraging a substantial 21-year dataset and a robust year-based holdout validation strategy with a diverse set of metrics. The findings are insightful, clearly highlighting the distinct strengths and weaknesses of each method under at-site and regional frameworks, and the operational recommendations provided are highly valuable for both researchers and practitioners in renewable energy integration. The paper is generally strong and merits publication after minor revisions to address the following points.

• Eq.(11) appears to contain a typographical error; the symbol for the mean value of y should not carry the subscript 'i', as it represents a single value summarizing the entire dataset or subset, not an individual observation.
• To improve the immediate visual impact and accessibility of the key results, I strongly recommend supplementingTable 2 with a statistical graph, such as a bar chart or a Taylor diagram. A visual representation would allow readers to instantly grasp the comparative performance differences.
• The introduction would benefit from a more nuanced and contemporary categorization of wind speed prediction methods. The current description could be expanded to more accurately reflect the broader taxonomy in the field, which typically includes (i) physical models (e.g., NWP), (ii) statistical models, (iii) artificial intelligence models, and (iv) hybrid models that combine these approaches.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Overview
The manuscript compares three probabilistic wind-forecasting methods across individual stations and a regional setting. The topic is relevant and the results are useful, but the text is overly long, contains many unnecessary details, and some parts repeat the same ideas. The structure is not very clear, and key points are harder to see than they should be. Because of this, the manuscript requires substantial shortening and reorganisation. Specific comments for each section are provided below.

The Title
The title is too long because it includes too many elements. A shorter title would be clearer, would express the main focus more effectively. 

Abstract
The abstract gives a clear overview of the main results and shows how the three forecasting methods behave under different modelling setups.
However, 
 - The abstract is too long.
 - There are too many numbers, which makes it hard to see what is most important.
 - It contains technical details that should be in the results section.
 - Some sentences are heavy and not easy to read.
 - The main idea of the study is not clearly highlighted. The aim is not stated in a simple, direct way.
 - The abstract includes more interpretation than needed.
Recommendation: The abstract should be shorter, clearer, and focused only on the main aim and key findings.

Keywords
The keywords are too detailed and include too many specific model names, which reduces their clarity and usefulness for indexing.

Introduction
The introduction gives useful background. However, 
 - it is too long and should be shortened by removing information that is not essential (detailed Australian energy policy figures);
 - some parts describe the models in too much detail (for example, long summaries of BART, GPR, QRF);
 - meteorological explanations are too long (e.g., extended descriptions of ECLs and ENSO) and should be reduced;
 - many sentences are too long and would be clearer if rewritten more simply (for example, multi-line descriptions of wind regimes);
 - the research gap comes only after several unrelated sections. It appears too late and should be stated earlier.
 - the aim and objectives are clear but expressed in unnecessarily long sentences (e.g., the aim paragraph spans several lines).

Section Study cite
The section gives a useful overview of the study region and the data used.
However,
 - the text is too long, and several details should be removed (for example, long descriptions of local wind regimes);
 - some parts repeat what is already shown in tables and figures (restating values that are visible in Table 1);
 - some parts include interpretation that belongs to the results  or discussion (for example, comments about tail risk or model behaviour);
 - the meteorological background is described in too much depth (for example, extended notes on ECLs or sea-breeze patterns);
 - many sentences are very long and could be made clearer ;
 - the summary statistics include more commentary than needed. For example, long explanations of precipitation skewness could be useful in the Discussion section).

Methodology
The methodology section presents the main steps of the modelling process clearly.
However,
 - it is too long. Some theoretical details could be reduced (for example, full BART and GPR formula explanations);
 - some parts sound more like teaching material than a method description (for example, long notes on kernel behaviour in GPR);
 - some sentences include interpretation that belongs in the results, for example, comments about handling tail events in QRF - it is an explanation not about "how", but about "why".
 - many sentences are very long and should be made shorter (e.g., extended descriptions of priors and hyperparameters);
 - the hyperparameter lists are too detailed for this section (e.g., full sets of tuning values for QRF, BART, GPR);
 - certain points are repeated

Results
The results section provides a broad and informative comparison of the models across sites, metrics, and forecasting regimes. Overall, the results section is the strongest part of the study. It clearly shows how each algorithm behaves, highlights their strengths and weaknesses, and links numerical findings to practical interpretation. 
Several points should be improved:
 - Some explanations are too long. For example,  extended reasoning about why GPR smooths coastal signals; long descriptions of pooling effects).
 - Certain numerical details could be reduced in the text. For example, precise shifts like “+0.006 m/s” or “+39.5%” - these are already clear from tables.
 - Some points are repeated across the text, for example, QRF RMSE stability, BART coverage drop, repeated notes about GPR coverage and QRF under-coverage

Conclusions
The conclusions outline the main differences between the models, but there are several issues :
 - Key findings are not highlighted clearly because the text is overloaded with secondary details.
 - The conclusions are too long and should be shortened to a clear summary.
 - Too many numbers and technical details appear; they belong in the results.
 - Several sentences are complicated and should be rewritten more simply.

Figures
The figures are not sharp. The resolution for Figures 5 and 8 should be increased. All axis labels should use a bigger, darker font (black). 

Comments on the Quality of English Language

The English should be improved.
Many sentences are unnecessarily long.
Some wording makes the main ideas difficult to follow.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript presents a comprehensive and well-structured comparative analysis of probabilistic wind speed forecasting methods under at-site and regional frameworks. The topic is timely and relevant; however, several major and minor issues should be addressed to further strengthen the methodological clarity, interpretation, and practical relevance of the study.

1. The Introduction would benefit from a clearer separation between background information and the specific research gap. Streamlining the narrative to more explicitly highlight the study’s novelty would improve clarity.
2. The severe degradation of BART coverage under regional pooling is an important finding. The link between this behavior and the use of fixed noise priors should be more clearly explained.
3. The operational meaning of the CRPS–RMSE trade-off should be clarified. A brief explanation of why GPR may still be preferred despite higher RMSE would strengthen the discussion.
4. While the chosen predictors are reasonable, a brief justification for excluding variables such as wind direction or large-scale climate indices should be provided.
5. Given potential climate-driven nonstationarity over 2000–2020, the adequacy of using “Year” as a proxy should be more clearly discussed.
6. The extent to which the results can be generalized beyond NSW and QLD should be briefly acknowledged as a limitation.
7. It is strongly recommended that a brief comparison of the calculation costs, especially for GPR, be included.
8. In Equation (13), the summation symbols and indices used in the denominator of the correlation coefficient are inconsistent with the rest of the manuscript and should be corrected.
9. Figure 2 is introduced in Section 2.4 but is not explicitly linked to the discussion of results; a direct interpretative reference should be added.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

I have completed my review of your revised manuscript titled "Probabilistic Wind Speed Forecasting under At-Site and Regional Frameworks: A Comparative Evaluation of Bayesian Additive Regression Trees, Gaussian Process Regression, and Quantile Random Forests."

I am pleased to inform you that you have successfully addressed all of my previous comments and concerns. The modifications you have made to the manuscript are satisfactory, and I have no additional comments or suggestions at this time.

Best regards,

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Overview
The paper compares three probabilistic wind-forecasting methods at individual stations and at a regional scale. The topic is relevant, the methods are sound, and the results are clear. The revised version is easier to read, with fewer repetitions and clearer main points. Overall, it is a solid study.

Some inaccuracies
 - Axes should include the variable name and units (e.g., observed and predicted wind speed in m/s)
 - The values in the tables use different decimal precision. Please consider using a consistent format.
 - The x-axis labels in some graphs use both “relative importance” and “importance”. Please use a consistent label.
 - The notation for r and R² is formatted inconsistently (italic vs. regular). Please standardize the formatting throughout the manuscript.
 - The reference list has not been corrected. The issues noted previously (duplicates, weakly related sources, inconsistent formatting, and incomplete entries) remain.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf