Tyan-WP: A Wind Power Foundation Model for Ultra-Short-Term Probabilistic Forecasting

Researchers introduce Tyan-WP, a foundation model for wind power forecasting pretrained on 126,000 U.S. sites that achieves superior accuracy without site-specific training. The model addresses critical challenges in renewable energy deployment by enabling rapid turbine onboarding and probabilistic risk assessment for new wind farms.

Tyan-WP represents a significant advancement in applying foundation model architectures to renewable energy infrastructure, addressing a genuine operational bottleneck in wind farm deployment. Traditional site-specific time series models require extensive historical data that new installations lack, while generic large time series models fail to capture the unique relationship between meteorological conditions and power generation at specific locations. The model's architecture innovatively combines static site metadata (coordinates, terrain, ecoregion data) with dynamic meteorological inputs through a power-aware fusion module, enabling zero-shot forecasting that outperforms both approaches.

The research validates a critical trend: foundation models trained on large heterogeneous datasets can transfer knowledge across geographies when properly designed for domain specifics. By training on seven years of data across 126,000 U.S. sites and successfully generalizing to U.K. locations, Tyan-WP demonstrates cross-border applicability that challenges the assumption that wind forecasting requires purely local models.

For the renewable energy industry, this has immediate practical implications. New wind farms currently face 3-6 month commissioning delays partly due to insufficient baseline data for grid integration and risk management. Accurate probabilistic forecasting reduces this timeline and improves grid stability planning. Energy traders and grid operators gain better risk metrics (CRPS and AQL improvements of 22% and 21.7%) for operational decisions.

The development trajectory suggests foundation models will increasingly enable rapid deployment of distributed renewable infrastructure. Future work likely focuses on extending this approach to solar, hydroelectric, and multimodal energy systems, potentially reshaping how utilities deploy and optimize renewable portfolios globally.

• →Tyan-WP reduces wind power forecast error by 19.9% MAE and 16.6% RMSE compared to conventional site-specific models
• →The model enables accurate zero-shot forecasting for new wind farms without requiring historical on-site training data
• →Foundation model approach demonstrates strong cross-geography generalization from U.S. training data to U.K. test sites
• →Domain-specific architecture innovations including static site embeddings and power-meteorological fusion modules drive performance gains
• →Accurate probabilistic forecasting accelerates turbine commissioning timelines and improves grid risk management capabilities