the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 21 Aug 2026
Influence of atmospheric stability on the wind speed-energy-revenue forecasting chain
Abstract. Ultra-short-term wind forecasting (< 30 min) is essential for the reliable operation and economic planning of wind energy systems, particularly in regions with high atmospheric variability. Atmospheric stability influences wind dynamics and predictive accuracy, but its role in propagating errors along the wind-energy-revenue chain has received limited attention. To address this gap, the relationship is qualitatively evaluated at a tropical site on the Yucatán Peninsula using 10-minute SoDAR measurements collected over a year of climatic variability. Representative forecasting approaches with different learning mechanisms, including persistence, support vector regression (SVR), and light gradient boosting machine (LGBM), were assessed across various forecast horizons, lag structures, and predictor configurations that included turbulence-related variables. The findings indicate that atmospheric stability significantly influences both prediction performance and the consequences of prediction errors. LGBM consistently outperforms both persistence and SVR, particularly in multi-step forecasting scenarios, and exhibits low sensitivity to lags and predictors. In contrast, SVR shows a stronger dependence on model configuration and improves when turbulent variables (turbulence intensity, turbulent kinetic energy, and vertical velocity) are incorporated, especially under stable conditions. Propagating forecast errors through a representative power curve and actual market prices shows that small biases near rated power can cause revenue deviations of 2.3–4.8 %. These results provide a systematic assessment of how atmospheric stability shapes forecasting skill and the operational and economic risk of ultra-short-term wind power forecasting.
- Preprint
(11683 KB) - Metadata XML
-
Supplement
(4707 KB) - BibTeX
- EndNote
Status: open (extended)
-
RC1: 'Comment on wes-2026-122', Anonymous Referee #1, 05 Oct 2026
reply
-
AC1: 'Reply on RC1', Mauricio Escalante, 26 Oct 2026
reply
We thank the reviewer for the careful assessment of our manuscript and for recognizing the relevance of investigating how wind speed forecast errors propagate to power and revenue estimates. We acknowledge that the manuscript requires substantial restructuring and clearer positioning of its contribution. Below, we provide preliminary responses and outline the revisions we propose to undertake.
Major comments
1. Electricity price consistency, horizon alignment, and decision-making context
We thank the reviewer for identifying the need to clarify the electricity price source, its temporal alignment with the forecasts, and the practical interpretation of the economic assessment.
The economic analysis reported in the manuscript uses Real-Time Market (MTR) local marginal prices obtained from CENACE for the Tizimín interconnection node, 08TZM-115. The hyperlink in the manuscript inadvertently directed readers to the Day-Ahead Market (MDA) portal instead of the relevant real-time price source. We apologize for this ambiguity. In the revised manuscript, we will correct the reference, explicitly identify the market and price series we used, and provide the corresponding mathematical formulations. We selected Node 08TZM-115 as a grid reference point in an area targeted for wind energy expansion in the Yucatán Peninsula.
We will also distinguish between the forecasting horizon and the temporal resolution of the price series. The models generate forecasts at 10-minute (t+1) and 30-minute (t+3) horizons, whereas the electricity prices used in the analysis have an hourly resolution. For the economic assessment, we aggregated the forecast series over six consecutive 10-minute intervals to align the resulting estimates with the hourly price signal. We will revise Section 3.6 to describe this aggregation procedure explicitly.
We agree with the reviewer that forecasts at horizons of 30 minutes or less are relevant to real-time operational adjustments and deviations rather than to schedules established in the day-ahead market. We will therefore clarify that the assessment uses MTR prices to value forecast-related deviations under real-time market conditions (real-time financial penalties, imbalance deviations, and spot market exposure experienced by wind power producers), rather than to represent day-ahead revenue.
To place the prices observed during the 11-day test period in a broader context, we will include a historical price analysis for Node 08TZM-115 covering 2022–2024, as presented in the accompanying figure. This analysis shows a recurring increase in prices and volatility during the evening ramp, beginning around 17:00 and peaking between approximately 18:00 and 21:00. The historical analysis will provide context for the economic assessment; it does not extend the meteorological evaluation period.
We will also clarify the intended relevance of the assessment for stakeholders. Our objective is neither to forecast electricity prices nor to simulate commercial strategies, but to quantify how wind speed forecast errors translate into power and revenue deviations under different atmospheric stability regimes. The results may inform plant operators, asset managers, and grid operators by:
- Providing an economic criterion for comparing forecasting models, particularly during periods of high price volatility.
- Identifying periods of elevated physical and financial exposure that may inform reserve allocation and regime-sensitive operational decisions during transitions to stable conditions.
Since the study does not simulate bidding strategies, storage dispatch, or reserve procurement, we will explicitly avoid claiming demonstrated improvements in profitability or operational performance.
2. Structure, narrative, and terminology consistency.
We agree that the manuscript needs a clearer narrative, more consistent terminology, and figures and tables that can be understood independently of the main text. We propose the following revisions:
- Terminology: We will standardize terminology throughout the text, figures, and tables, using “SVR” rather than alternating between SVR and SVM, and “LightGBM” rather than alternating between LightGBM, LGBM, and GBM. We will also standardize the terminology and notation for wind shear.
- Abbreviations: We will review all abbreviations and symbols, define them at first use, and reduce unnecessary abbreviations. We will consider including a nomenclature table in the appendix.
- Measurement period: We will correct the description of the measurement period in the abstract so that it is consistent with the main text.
- Captions: We will rewrite the figure and table captions to make them self-contained. For example, the caption of Figure 8 will identify the model and the method used to calculate feature importance, while Figure 12 and its axis labels will explicitly identify the Pasquill–Gifford (PG) stability classes.
- Organization and presentation: We will review the manuscript to remove redundant material and improve the sequence of the methods and results. Secondary evaluations, detailed parameter configuration tables, and supplementary correlation plots will be moved to an appendix or supplementary material.
These changes will focus the main text on the study’s central contribution: the propagation of wind speed forecast errors to power and revenue estimates under different atmospheric stability conditions.
3. Limited evaluation set size (11 days), statistical confidence and significance
We appreciate the reviewer’s concern regarding the 11-day test period, particularly when the observations are further divided into atmospheric stability classes.
We divided the dataset chronologically into training, validation, and test subsets comprising approximately 80%, 13%, and 7% of the observations, respectively. The larger validation subset supported hyperparameter selection, while the final test block was reserved for out-of-sample performance assessment.
It must be highlighted that the evaluation did not rely on a single forecast origin. We applied a fixed-length sliding-window backtesting procedure in which the oldest observations were removed and newly available observations were incorporated at each successive forecast origin, followed by model retraining. This allowed performance to be assessed across successive out-of-sample forecasts within the test period.
Nevertheless, we acknowledge that repeated evaluation within an 11-day period does not establish representativeness across a broader range of meteorological conditions, and stratification by atmospheric stability further reduces the sample size within each regime. However, this evaluation window captured two extreme synoptic events: a hurricane and a cold front. The rapid boundary-layer transitions and turbulence generated by these phenomena served as a stress test under severe, non-stationary tropical conditions.
We will therefore qualify the interpretation of these results throughout the manuscript. Differences between stability classes will be presented as patterns observed during the evaluation period, rather than as statistically established or broadly generalizable differences. We will also clarify that atmospheric stability was used to characterize and stratify forecast errors, not as a prediction target or an input feature.
Evaluation over longer periods and under a wider range of meteorological conditions will be identified as an opportunity for future work.
Other remarks:
1. High volume of presented results
We agree that the presentation should prioritize the results that convey the main insights.
We will condense Section 5 to emphasize the principal model comparisons, forecast behavior during the afternoon transition, and the propagation of errors to revenue estimates. Secondary configuration tables and baseline diagnostic plots will be moved to an appendix or supplementary material.
We will also streamline Section 2 to retain only essential visualizations, moving extended explanations to the supplementary material.
2. Discrepancy in revenue impact (Persistence vs. ML models)
We thank the reviewer for highlighting this imbalance in the presentation.
We will revise the abstract and Section 4.4 to acknowledge explicitly that the persistence model exhibits smaller revenue deviations, despite producing larger wind speed errors at certain forecasting horizons. The revised discussion will explain this contrast and avoid presenting the revenue impact observed for particular models as a uniform finding across all forecasting approaches.
3. Inclusion of PG stability class in Correlation Matrix (Fig. 7)
We thank the reviewer for this suggestion.
The PG stability classes were used exclusively to stratify forecast performance by atmospheric regime, rather than as input features for model training. Nevertheless, we agree that including them in the correlation matrix first can provide useful physical context and later corroborate our findings.
We will update Figure 7 to include the PG class index alongside the predictor variables. In Section 4, we will explicitly state that its inclusion serves diagnostic and interpretative purposes only, illustrating its association with turbulence and wind shear variables without implying that PG classes were used as predictors.
4. Order of preprocessing steps and data leakage
We will revise Section 2.3 to describe the preprocessing sequence explicitly and include a flowchart showing the order of operations.
The revised description will confirm that the chronological data split preceded data imputation and feature scaling, making the separation between preprocessing and model evaluation clear.
5. Comparison with direct revenue/power forecasting
We thank the reviewer for this suggestion.
In the present study, power output and revenue are estimated using the manufacturer’s theoretical power curve and local marginal prices because turbine SCADA measurements are unavailable. Consequently, the target values for a direct supervised power or revenue model would themselves be derived quantities rather than independently measured operational outcomes.
Our choice of an indirect forecasting approach reflects the study’s objective: to trace wind speed forecast errors through the nonlinear power transformation and subsequent price-based valuation. This approach allows us to distinguish the role of the initial forecast error from the effects of the power curve and the timing of electricity prices.
We will add a paragraph to the methodology section explaining this rationale and clarifying why the indirect approach was selected instead of direct power or revenue forecasting.
6. Self-contained captions for figures and tables.
We agree that readers should be able to interpret the figures and tables without repeatedly consulting the main text.
We will systematically revise all captions to identify the relevant variables, models, forecasting horizons, measurement heights of 30 m and 80 m, and atmospheric stability classes, as applicable. Model-specific quantities and calculation methods will also be explained where needed.
-
AC1: 'Reply on RC1', Mauricio Escalante, 26 Oct 2026
reply
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 464 | 179 | 124 | 767 | 131 | 146 | 121 |
- HTML: 464
- PDF: 179
- XML: 124
- Total: 767
- Supplement: 131
- BibTeX: 146
- EndNote: 121
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors aim to assess the impact of atmospheric stability (metrics) on wind speed, power and revenue forecast errors. To this end, they develop short-term wind speed forecasting models using persistence, support vector regression and gradient boosting machines. They then apply a power curve and market prices to transform these errors.
Overall feedback: The paper requires substantial restructuring and clearer positioning before its contribution can be properly assessed.
Positive:
Major remarks:
Other remarks: