Presentations

Abstract

When wind power forecasts are not accurate, operators and market participants face high costs, imbalance penalties and challenges for system stability. Improving prediction is therefore key to make wind a more reliable source of energy. This work describes a forecasting approach that combines traditional Numerical Weather Prediction (NWP) models with machine learning (ML) techniques. NWP models provide valuable physical insights, but they often leave systematic errors. ML can learn from data such as past SCADA production, site topography, and turbine features to correct and adapt forecasts over time. Together, these methods provide more robust results than on their own. In addition, forecasts are produced across short-term horizons (minutes to 48 hours), essential for intraday and day-ahead markets, as well as medium- and long-term horizons (days to months) to support planning and asset management. The methodology is implemented in a scalable cloud-based environment, ensuring continuous availability and integration with grid operation and market systems. Having all these capabilities in one solution avoids fragmentation, ensures consistency, and makes the tool more practical for industry use. The system learns continuously from new SCADA data, turbine characteristics, and site information, correcting biases from physical models and adjusting to seasonal patterns. Tests on real wind farms show error reductions which is critical for market participation. By providing a single forecasting framework with multiple scales, this work supports better decision-making for both operations and planning, helping reduce costs and making wind energy a more predictable and reliable resource for the future energy system.