An innovation project between National Grid ESO and The Alan Turing Institute has harnessed machine learning to improve solar forecasting accuracy by 33 per cent.
The ESO has historically used a relatively simple way of forecasting, based on installed solar capacity and solar irradiance across different regions to create a national forecast. On the innovation project with Turing, the two combined to trial a novel, more complex approach.
The method involved a ‘random forest’ approach, taking in historic data and around 80 input variables and much more granular solar irradiation data.
Using machine learning, the random forest model trains itself by finding hundreds of different mathematical pathways (or decision trees) to take those inputs, and arrive at the output generation figure. The model then forecasts by running the 80 new forecast weather variables through these decision trees, and takes the average as the new solar generation forecast. The ESO took the new approach, and combined it with several other machine learning techniques in a multi-model ensemble forecast.
By combining the project’s initial output with these different machine learning methods, the ESO said it has built a solar forecasting system which is 33 per cent more accurate.
“Improved solar forecasts will help us run the system more efficiently, ultimately meaning lower bills for consumers. It will also enable more solar capacity to be connected and utilised, helping us to achieve our 2025 ambition to be able to operate a zero carbon electricity system,” said Rob Rome, Commercial Operations Manager at the ESO.
Andrew Duncan, Data-Centric Engineering Group Leader at The Alan Turing Institute added: “The project has opened up a lot of new avenues and ESO are interested in pushing other projects forward. There’s no shortage of problems to tackle”.
The project, funded via Ofgem’s Network Innovation Allowance, is one of a number that have enabled improved solar forecasting. Earlier this month the ESO published results of a project with the Met Office that have helped deliver improvements of between 5 and 9% for the in-day and day-ahead national demand forecasts over the last financial year.
Solar PV capacity currently stands at around 13GW, virtually all connected at distribution level, over which the system operator has limited visibility.
As such, accurately forecasting solar irradiation levels can have a significant impact on how the system operator manages to keep supply and demand in balance while maintaining frequency levels within set tolerances.