At a time when the energy transition is advancing steadily, improving the ability to forecast renewable energy generation has become a critical and highly valuable necessity.

Accurately predicting how much solar or wind energy will be produced in the coming hours or days enables key decisions to be made for the stability of the power system, market efficiency, and the integration of renewables.

In this context, Ravenwits, with support from the Community of Madrid, is making progress in developing new forecasting models based on artificial intelligence, as well as expanding their use to make them an accessible large-scale solution.

From traditional forecasting to applied AI

Until recently, renewable generation forecasting relied on traditional statistical methods combined with weather data. Ravenwits has introduced a more advanced approach: convolutional neural networks (CNNs), an AI technique particularly effective at identifying complex patterns.

This technique, applicable to both wind and solar energy, has already proven its effectiveness in various contexts. First, in a pilot project with Red Eléctrica de España that later gave rise to Ravenwits, and more recently in commercial solutions already used by leading players in the energy sector to forecast the generation of their plants.

Thanks to these AI-based forecasting models, wind and solar prediction accuracy has improved by approximately 5% compared to competing models. This improvement has been verified with real-world data and is already being applied daily by several clients.

However, the work is not yet complete. Despite the success of the current models, significant challenges remain: improving forecasts over short- and medium-term horizons (from hours to several days), solar nowcasting (predictions from a few minutes up to 1 or 2 hours), and automating the process to scale the solution across hundreds or thousands of plants.

New challenges, new tools

One of the main goals of the new project funded by the Community of Madrid is to explore new artificial intelligence techniques that can overcome current limitations. Specifically, the project will design models based on Graph Neural Networks (GNNs), an architecture designed to capture complex and non-local relationships between different geographical points.

Unlike CNNs, which excel at detecting local patterns in image-like data, GNNs enable the modeling of broader relationships—such as how wind fronts or clouds propagate between distant areas. This is key for improving forecasts with 24- to 48-hour horizons (which have the greatest impact on electricity auctions), especially in complex geographies.

With the rapid growth of installed photovoltaic capacity worldwide—particularly in Spain—a new critical challenge emerges: solar nowcasting. This term refers to forecasting photovoltaic generation in the very short term, typically for the next one or two hours.

In this case, traditional weather forecasts are not sufficient, as they involve several hours of delay. For this reason, Ravenwits is working with satellite imagery updated every few minutes and CNN-based models adapted to anticipate the appearance or disappearance of clouds that may affect real-time production.

Scaling to create impact

A critical aspect of the project is the scalability of forecasting solutions. Currently, the process of validating data and training models often requires manual intervention: reviewing production data for errors, checking that training was successful, adjusting parameters...

This approach is feasible when working with a limited number of plants, but when scaling to a national or international level, manual supervision becomes unsustainable. That’s why the project also includes the development of AI-based automated tools capable of validating data and models without constant human input.

This not only facilitates applying the solution to an increasing number of installations, but also helps transform it into a robust, reproducible, and accessible product for a broad range of users and stakeholders in the electricity system.