Graph Neural Networks: a new era in renewable energy forecasting

Forecasting the real-time output of a solar plant or wind farm remains one of the biggest challenges in the power system. Unlike conventional sources, renewable energy depends on weather conditions that change continuously and often in a non-uniform way across different locations.

This challenge becomes even more critical as the share of renewables in the energy mix increases. The higher the penetration of renewable generation, the greater the need for accurate and reliable forecasting.

Traditionally, the industry has relied on classical statistical models, such as linear regressions or autoregressive models, combining historical production data with weather forecasts to estimate renewable generation. While these methods have proven useful in systems with low or medium levels of renewables, they show limitations when higher spatial precision and adaptability to abrupt weather events are required.

In this context, artificial intelligence has become an essential ally. Over the past few years, many companies, including Ravenwits, have incorporated deep learning techniques such as Convolutional Neural Networks (CNNs).

Now, a new generation of models is pushing the boundaries even further: Graph Neural Networks (GNNs).

What is a GNN, and why can it improve current forecasting methods?

A Graph Neural Network is a type of neural network designed to work not with images or time series, but with data structures shaped as graphs, that is, networks of interconnected nodes. This architecture enables the analysis of complex relationships between elements, such as meteorological stations, solar plants, or wind farms distributed across different locations.

Unlike CNNs, which analyze data as fixed pixel grids, GNNs learn how nodes interact and influence one another. This is especially useful in weather and energy forecasting, where what happens in one geographic location can affect nearby areas and those spatial interdependencies can significantly improve forecasting accuracy.

One clear example of the potential of GNNs comes from DeepMind, Google’s AI division. In 2023, they published a GNN-based global weather forecasting model that outperformed traditional models like those of ECMWF (European Centre for Medium-Range Weather Forecasts) across many metrics.

This breakthrough has opened up promising new opportunities for the energy sector. GNNs are expected to better capture the atmospheric dynamics that influence renewable generation, making them more aligned with the distributed nature of renewable plants and the meteorological phenomena that affect them.

From CNNs to GNNs: how is Ravenwits applying them?

Until now, Ravenwits has successfully used CNN-based renewable forecasting models. These networks are particularly effective at identifying spatial patterns in weather maps produced by numerical weather prediction (NWP) models, such as cloud cover, radiation, or temperature. The logic is similar to a system that “reads” images.

However, CNNs have a limitation: they only perceive local relationships within a fixed frame. They do not inherently account for how distant points interact with each other.

This is where Graph Neural Networks come in. GNNs can model interactions between multiple meteorological variables across different locations, essentially creating a network of nodes (plants, stations, measurement points), where each node influences the others.

Imagine three solar plants aligned west to east: A, B, and C. A traditional model analyzes each plant’s local data in isolation. So, to forecast production at plant C, it only looks at irradiance, cloud cover, or temperature at that location. If a cloud approaches from the west, the model will not anticipate the drop in generation until the cloud is already over plant C.

A GNN, on the other hand, learns the relationships between plants. If it detects a drop in irradiance at A, and then at B, it can predict that plant C will be next, even before C’s own sensors show any changes.

Practical applications in wind and solar energy

To improve the accuracy of renewable energy forecasts over short- and medium-term horizons (from several hours to two or three days), Ravenwits is developing a new deep learning model with a GNN architecture, supported by the Community of Madrid through its 2024 Grant Program for AI in Industry.

Ravenwits is now adapting these models for energy sector applications, with a strong focus on medium-term forecasting. Initial results using historical data have been very promising. The next step is to refine and validate the models under real operational conditions, making them reliable and practical forecasting tools.

The current phase of the project focuses on wind power forecasting, combining historical generation data with wind forecasts at 100 meters height. The approach will later be extended to solar PV forecasting, where the greater complexity of meteorological variables (irradiance, cloud cover, temperature, aerosols, etc) requires adjusting the model architecture to handle this diversity of inputs.

In a landscape where forecasting accuracy is more critical than ever, Ravenwits has a clear objective: to raise the standard of commercial forecasting systems by applying advanced artificial intelligence techniques.