PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
March 15, 2026Energy and AI0 citationsOpen Access

Spatio-temporal Transformer-based modeling of multi-turbine wake effects for wind farm power forecasting

View Full Paper
AAAdel AlaeddiniGEGustavo Alberto Giraldo EcheverriRARahmat Adesunkanmi

Key Points

  • The aim is to enhance short-term wind power forecasting in dense wind farms affected by turbine wake interactions.
  • Developed a spatio-temporal Transformer model for turbine-level forecasting.
  • Ingested operational power and geometric descriptors for accurate predictions.
  • Evaluated on data from an 86-turbine wind farm with 10-minute intervals.
  • Achieved a mean absolute error (MAE) of approximately 0.072918 MW.
  • Improved forecasting accuracy by about 3.7% compared to machine learning baselines.
  • Revealed systematic power-loss patterns associated with turbine spacing and angle.

Abstract

Short-term wind power forecasting is vital to reserve scheduling and reliable grid operation. However, accuracy degrades in dense wind farms where turbine wake interactions introduce strong, direction-dependent spatial coupling. We present a lightweight spatio-temporal Transformer for turbine-level forecasting that explicitly learns wake superposition in a multi-source ( N -to-1) setting, where several upstream machines jointly influence a downstream target. For each upstream-downstream group, the model ingests operational power along with physically meaningful geometric descriptors (inter-turbine distance and encoded wind-referenced bearing and angle difference) to preserve directional periodicity. The model performs one-step-ahead ( τ = 1 ) prediction at a 10-minute resolution for all G = 45 downstream turbine groups and is evaluated on 1 year of 10-minute-interval data from an operating 86-turbine wind farm. Using 45 interaction groups with a practical 5-to-1 neighborhood instantiation, the proposed model delivers lower errors than representative machine learning baselines, achieving an MAE of ≈ 0 . 072918 MW during the peak-wake evaluation period and improving MAE by ≈ 3 . 7 % relative to the best baseline. Beyond accuracy, the learned representations support wake-effect interpretation and reveal systematic power-loss patterns with turbine spacing and wind-aligned angular configuration, informing layout assessment and operational monitoring. • Spatio-temporal Transformer-based forecasts wind-farm power from space–time SCADA data. • Multi-source wake effects modeled with turbine spacing, angles and wind direction. • Transformer cuts mean error by ≈ 45% vs FFN, LSTM, TCN, and MAD-PCA. • Key wake-loss drivers: turbine spacing, alignment, and diurnal stability. • Study offers guidance for cost-effective wind-farm design and wake-loss mitigation.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Alaeddini et al. (2026) studied this question.

synapsesocial.com/papers/69b606af83145bc643d1ce49https://doi.org/10.1016/j.egyai.2026.100715
Ask AI
Helpful
Bookmark
Share
View Full Paper