In today’s energy transition landscape, the installed capacity of solar photovoltaic power is growing rapidly worldwide. In countries with high solar penetration, such as Spain, this shift in the energy mix brings a new challenge: managing the variability of solar generation on very short timescales. This is where nowcasting comes into play.

The term nowcasting combines the words now and forecasting, and refers to the prediction of events in the very short term, typically over the next minutes or a few hours.In the power sector, solar nowcasting means forecasting PV generation over time horizons from a few minutes up to two hours.

This time window is critical for both grid operators and power plants, as it allows them to anticipate sudden drops or spikes in output, often caused by the passage of dense clouds or weather fronts.

Unlike traditional numerical weather prediction (NWP) models, which have low temporal resolution and often include delays of several hours, nowcasting demands more immediate and accurate data.

The need for precise nowcasting services has become increasingly important as solar generation saturates the grid during peak hours. Decisions such as whether to feed energy into the grid, store it, or activate backup sources must be made quickly. This makes nowcasting a strategic tool for both grid operators and asset managers.

Looking at the sky: the role of satellite imagery and CNNs

But how can we forecast the solar output of a specific plant minutes or hours ahead? To anticipate short-term changes, the key tool is no longer numerical models, but satellite imagery, which allows us to monitor the atmosphere in near real time.

Institutions like EUMETSAT provide images every five minutes. This enables precise tracking of cloud cover movement and density over solar plants. These images offer a near-instantaneous snapshot of atmospheric conditions.

However, interpreting these images requires deep learning models. In this field, Convolutional Neural Networks (CNNs) have become the most effective solution. With their ability to “read” weather maps as if they were images, CNNs can detect cloud evolution patterns, predict their trajectory, and estimate their impact first on solar irradiance, and then on PV generation.

Building on its experience in renewable generation forecasting using CNNs, Ravenwits, within the framework of the Community of Madrid’s grant program for AI in industry, has adapted this technology specifically for the nowcasting challenge. The Madrid-based startup has developed a system that combines satellite images, minute-level solar production data, and weather forecasts to train CNN architectures capable of predicting solar generation up to two hours in advance.

The project includes tasks such as automatic downloading of satellite data, integration with historical production records, and validation of different CNN models in five-minute blocks. Additionally, the team is evaluating the integration of Graph Neural Networks (GNNs) to capture spatial relationships between nearby plants.

Benefits for the grid and power producers

Accurately forecasting changes in solar output at minute-scale resolution has a direct impact on grid stability.For grid operators, having access to reliable short-term forecasts enables better dispatch planning and reduces the need for activating backup generation reactively. During periods of high PV penetration, this lowers the risk of curtailment and improves system-wide efficiency.

For solar producers, nowcasting offers a competitive edge. Being able to anticipate generation drops allows them to activate storage, store energy, or adjust their market strategy. In a context where flexibility is becoming increasingly valuable, real-time forecasting is a key tool for both asset management and profitability.

In short, using artificial intelligence to “look at the clouds” is no longer a future possibility. It is a present necessity. Nowcasting is emerging as a crucial tool for managing renewable generation. Technologies such as CNNs and GNNs will be essential in shifting from reactive management to truly predictive operation.