Abstract
This paper describes the WindAI data challenge on short-term (48 hours ahead) wind power forecasting in Norway. The competition provided multi-year open datasets combining ensemble weather forecasts, nowcasts, aggregated wind power production, and wind park metadata for Norwegian bidding zones. Participants were tasked with producing day-ahead hourly forecasts under realistic operational constraints. We summarize the data, task design, evaluation protocol, and the modeling approaches adopted by the top-performing teams, and discuss key findings regarding model complexity, feature engineering, and robustness.