From foundation weather models to renewable operations: Adapting the WeatherGenerator for wind and solar production forecasting

The rapid growth of wind and solar power is transforming energy markets, but their inherent variability makes accurate, real-time forecasting more essential than ever. Errors in day-ahead forecasting directly drive up imbalance costs, while the fast-paced nature of intra-day trading requires model inference that is much faster than traditional weather simulations. Foundation weather models such as the WeatherGenerator (WG) offer strong generalization and the potential for low-latency deployment, but their value for the energy sector depends on effective adaptation, as they are not originally designed for plant-specific tasks.

We will present results from applying WG to site-level wind and solar production forecasting in Turkey. The downstream task targets individual plants and is trained and evaluated on historical production observations across a multi-site portfolio. Our focus is on adapting WG for this operational setting by evaluating a spectrum of adaptation strategies, ranging from training task-specific 'tail' networks to fine-tuning the entire model. We report how these choices affect forecast performance and consistency across different sites and conditions, and we describe the resulting workflow in a form that can be carried over to portfolio-scale deployment.

Performance is benchmarked against our current operational baseline, which combines NWP results with machine-learning post-processing. We report MAE as the primary metric and discuss application-oriented indicators that relate forecast improvements to operational value in day-ahead and intra-day settings. The goal is to provide practical guidance on how to translate a foundation weather model into measurable benefits for renewable energy forecasting workflows.