Predicting Wind Turbine Power Using Machine Learning and Weather Forecasts
Khivishta Boodhoo, Isaac Triguero, Josh Plumbly, Bruce Nicolson, Nicholas Watson
cs.AI · cs.LG
Quick verdict
ANN forecasts 7MW turbine power with R2=0.98 from weather data, but shuffled time-series splits risk inflated test metrics; deployment MAE is 7.5x higher.
Summary
This case study develops machine-learning models to predict the power output of the 7 MW Levenmouth offshore demonstration turbine (Scotland) from weather variables, so maintenance can be scheduled in low-output 4-hour windows up to 72 hours ahead. Using 2019 SCADA and met-mast data (~30 million 1-second samples, reduced to 48,956 four-minute rows), the authors apply startup/shutdown trimming, brake-state and negative-power filtering, Random-Forest feature selection (wind speed, direction, temperature, pressure), Mahalanobis-distance outlier removal, and min-max normalization. A Linear Regression baseline (wind speed only) is compared with a feedforward, grid-search-tuned ANN on randomly shuffled 60:20:20 splits. On one year of data the ANN reaches R2=0.98, MAE=194 kW and NRMSE 11%, versus the baseline's R2=0.94 and MAE=441 kW; preprocessing alone lifts the baseline from R2=0.31 to 0.94. Deployed on commercial forecasts from a station about 14.5 miles away (Jan–Mar 2022), error rises to MAE 1462 kW. A worked example suggests ~2000 kW of average power saved per maintenance event. Main limitations: shuffling a time series before splitting risks temporal leakage that inflates the headline metrics, the baseline uses fewer inputs, only one turbine is tested, and the forecast dataset is proprietary.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A feedforward ANN trained on one year of preprocessed SCADA data achieves R2=0.98, MAE=194 kW and NRMSE 11% on the shuffled test set, versus R2=0.94 and MAE=441 kW for the linear-regression baseline (Table 6).
- Data preprocessing alone lifts the LR baseline from R2=0.31 to R2=0.94 on one year of data (Table 6), demonstrating preprocessing is a first-order factor.
- Increasing training data from one month to one year improves both models (LR R2 0.81 to 0.94; ANN 0.94 to 0.98; Tables 5-6).
- Deployed on weather forecasts from a station ~14.5 miles away, the ANN predicts Jan-Mar 2022 power with MAE 1462.17 kW, 610.83 kW better than the baseline (Table 6, Conclusion).
- Scheduling 4-h maintenance in predicted low-output windows saves approximately 2000 kW of average power per event, based on a single worked example from October 2019 (Fig. 8).
Strengths
Honest deployment-error reporting
Unlike much of the literature it critiques, Tables 5-6 explicitly report the error between SCADA actuals and predictions driven by the independent METEO forecast source (e.g., ANN deployment MAE 1462.17 kW vs LR 2072.99 kW for one year of data), alongside the much flatter test-set metrics.
Preprocessing impact is cleanly isolated
Table 6 shows the LR baseline moving from R2=0.31 (no preprocessing) to R2=0.94 (full preprocessing) on the same one-year data, and Table 3 quantifies that 99% of negative-power points are attributable to applied brake states - a genuinely instructive ablation.
Real industrial-scale data with a public training set
End-user framing is concrete
Weaknesses
Temporal leakage in the headline evaluation
Baseline comparison is confounded
Deployment gap left unexplained
Ad hoc preprocessing rules lack validation
Internal inconsistencies suggest limited proofreading
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This Energy and AI / arXiv paper (Boodhoo, Triguero, Plumbly, Nicolson, Watson) is a single-turbine case study: predict the 7 MW Levenmouth offshore turbine's output from weather variables only, so maintenance can be scheduled in low-output 4-h windows up to 72 h ahead. Five claims anchor the paper: (1) an ANN beats an LR baseline (R2=0.98 vs 0.94, Table 6); (2) preprocessing is first-order (LR 0.31 to 0.94, Table 6); (3) data volume matters (month vs year, Tables 5-6); (4) weather-station-only inputs enable deployment on future forecast data (deployment MAE 1462 kW, Jan-Mar 2022); (5) low-window scheduling saves ~2000 kW per maintenance event (Fig. 8).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How accurate is the wind turbine power prediction model in this paper?
On a randomly shuffled test set built from one year of SCADA data of the 7 MW Levenmouth turbine, the ANN reached R2=0.98 and MAE=194 kW (NRMSE 11%), versus R2=0.94 and MAE=441 kW for the linear-regression baseline. When driven by independent weather forecasts from a station about 14.5 miles away (Jan-Mar 2022), the error rose to MAE 1462 kW - a useful reminder that test-set metrics overstate real forecast skill.
What data did the study use and is it available?
Two sources: (1) 2019 SCADA and met-mast data from the 7 MW Levenmouth demonstration turbine in Scotland, about 30 million one-second samples, publicly available via the ORE Catapult POD; (2) commercial METEO weather forecasts from the Intelligent Plant Appstore, taken from a MET station roughly 14.5 miles from the turbine and used to simulate deployment for Jan-Mar 2022. The forecast dataset is proprietary; no code is released.
How does the model help schedule wind turbine maintenance?
Predicted power is aggregated into 4-hour blocks up to 72 hours ahead so crews can pick low-output windows for shutdowns. In one worked example from the SCADA records, rescheduling a 4-hour maintenance visit from a ~5126 kW average-output window to a ~3000 kW window would have saved about 2000 kW of average power - though this is a single illustrative event, not a fleet-wide statistic.
Can this approach be transferred to other wind turbines?
In principle yes: the model inputs are common weather variables (wind speed, direction, air temperature, pressure) that can come from a weather station rather than turbine sensors. However, the paper tests only one turbine with one nearby station, so cross-turbine and cross-site performance remains unverified, and several preprocessing rules are site-specific and ad hoc.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are image content not present in the text extraction; judgments about Figs. 7-10 rely on captions and in-text descriptions only.
- Supplementary Table 1 and supplementary figures are referenced but not included in the provided text, so the cross-study comparison claims could not be independently checked.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.