Solar field with glowlines

Solar NowCasting innovation project improves solar forecasting

Following the launch of our latest Innovation Annual Summary, we’re taking a deeper dive into our Solar NowCasting project, in collaboration with Open Climate Fix, exploring how we can increase the amount of solar generation the grid can handle as we continue our journey to operating a zero-carbon electricity system.

What is Solar NowCasting?

As solar generation on the electricity network continues to increase, it brings with it a number of challenges. Firstly, solar is embedded generation, meaning it is connected to a distribution network rather than the transmission network, making it invisible to grid operators. Secondly, existing weather data doesn’t give enough information to allow operators to accurately predict generation from known solar resources, like large scale solar farms. As a result of this uncertainty, our Control Room must keep power in reserve to make up any shortfalls in solar generation caused by cloud cover or unexpected changes in the weather, particularly in the ‘shoulder months’ of April/May and September/October when the weather is more unpredictable.

This means we must bolster solar electricity generation with fossil-fuel-powered generators, which are typically carbon-intensive and more expensive than renewables.

Using deep machine learning techniques, our Solar NowCasting project is exploring how more accurate predictions for solar electricity generation could reduce the amount of power we need to keep in reserve to cover unexpected shortfalls. This would reduce carbon emissions and reduce costs to end-users, as well as increase the amount of solar generation the grid can access.

How can deep machine learning improve our solar forecasting?

The project has developed a deep learning model, which takes a sequence of recent satellite images and numerical weather predictions. The model develops probabilistic solar electricity nowcasts for each solar cell system across Great Britain, which will be calibrated in near-real-time using live solar electricity data. Deep learning models can handle huge amounts of data, so we can train the model across the entire geographical extent of the satellite imagery (not just the areas that happen to have solar electricity systems). As such, the model can predict the next few frames of satellite imagery as well as solar electricity generation.

Improving our solar forecasting

The national solar generation forecast developed during this project was 2.8 times better than our previous Photo Voltaic (PV) forecast (for forecasts up to two hours ahead). The best national PV forecasts developed to date have a mean absolute error (MAE) of 233MW, compared to our existing national solar PV forecasts, which have an MAE of 650MW. 

The first version of a fully operational PV Nowcasting service was delivered to our control room in December 2022. As part of the further work extension, the project is now investigating the addition of further features and improving the accuracy of the forecasting algorithm by at least 20%.

Find out more

Learn more about each stage of this project on the ENA Smarter Networks Portal.

You can also read our Innovation Annual Summary to find out how our innovation projects are essential for the transition towards a sustainable energy system, enabling us to execute ambitious projects and drive collaboration across the whole energy sector.

Innovation Annual Summary