Renewable energy forecasting with Deep Learning tools

PV forecasts for both the mainland and the island systems. These forecasts, updated hourly and extending up to 15 days ahead, are integrated into control center processes and published for market participants and the public. To enhance robustness, Red Eléctrica operates a cloud-based forecasting platform that combines proprietary and third-party models in a multi-model ensemble.

Recent advances in data-driven approaches, particularly Deep Learning, have opened new possibilities for improving forecast accuracy. In 2020, the Spanish TSO launched a multi-year

R&D initiative with academic partners to assess the applicability of Deep Learning to RES forecasting. The project began with proof-of-concept (PoC) wind-forecasting models for peninsular Spain using 3D Convolutional Neural Networks (CNNs), yielding promising results that led to continued collaboration and the development of dedicated architectures for both wind and PV forecasting.

The resulting models include system-level forecasters for the Spanish mainland and Canary

Islands, as well as high-resolution models for each wind installation (~1500 plants) and each

PV plant above 5 MW (~800 plants). These granular forecasts support real-time operation and short-term network scenario generation. Model inputs combine historical production data with

Numerical Weather Prediction (NWP) fields, complemented by the latest available production for very short-term horizons.

Key insights from the project highlight the importance of spatial context: wind forecasting benefits from input regions spanning at least 200 km to capture mesoscale patterns, whereas

PV forecasting performs best with smaller, localized spatial patches. While CNNs currently lead in wind-forecasting performance, decision-tree models have proven highly competitive for

PV due to their speed, robustness with limited data, and strong predictive capability.

In June 2025, the Spanish TSO deployed the new wind-forecasting models in production after confirming significant accuracy improvements, including a 1.58% reduction in forecasting error during a 2022 evaluation period and a 3.5% improvement in May 2025. Upcoming work includes online validation of individual wind-farm and PV models, with initial offline results showing strong potential. The forthcoming article will provide detailed performance analyses and lessons learned.