3.5 A Machine Learning Based Approach to Predicting Probabilistic Power Generation at Individual Wind Farms in ERCOT

Monday, 29 January 2024: 2:45 PM
347/348 (The Baltimore Convention Center)
Maxfield E Green, MA, Tomorrow.io, Fort Collins, CO; and T. McCandless, L. Conibear, B. Taylor, A. E. Payne, A. Reed Harris, K. Keshavamurthy, and S. Flampouris

The wind power forecasting solution presented here provides utilities and wind farm operators with critical hourly predictions up to 48 hours in advance, enabling highly informed operational and trading decisions. By utilizing the High Resolution Rapid Refresh (HRRR) model and processing data from multiple vertical levels, we developed a machine learning-based model to predict each wind generating resource in the Electric Reliability Council of Texas (ERCOT). Our machine learning model was trained on 4-years of historical wind power generation data from the ERCOT, and results for this domain show highly accurate site-specific deterministic and probabilistic forecasts with R2 of 0.94 for hours 1-24 at the regional level, and an R2 of 0.91 for hours 25-48. The machine learning model is operational in real-time and supports decision-making in the day-ahead and real-time markets.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner