- 1Open Climate Fix, London, United Kingdom, (djamesfulton@yahoo.co.uk)
- 2The Alan Turing Institute, London, United Kingdom
- 3Queen Mary University, London, United Kingdom
The global transition to low or no carbon electricity grids requires the use of a large amount of renewable energy sources such as photovoltaic solar power. However, to integrate these intermittent energy sources within stable electricity grids requires accurate solar power generation forecasts.
Satellite imagery is highly valuable for making short-term forecasts of solar generation. The stream of satellite data is low latency, usually only minutes behind real-time, is measured frequently, and is a direct measurement of the atmosphere. This complements numerical weather predictions (NWPs) which take several hours to compute from initial conditions, generally produce forecasts at only hourly steps, and are simulated and so have an imperfect and limited expression of the atmosphere.
Including satellite data often makes for better solar forecasts than using NWPs alone. However, for solar forecasts at time horizons beyond a couple of hours, satellite imagery becomes less and less useful as the atmospheric conditions will continue to evolve beyond those captured in the most recently available satellite image.
In this work, we introduce a machine learning model to forecast upcoming satellite images from recent satellite images. This can be done using relatively simple neural network architectures designed for video prediction. We show that we can increase the accuracy of solar generation forecasts in Great Britain by using these forecasted satellite images instead of just using recent satellite images.
We find that using predicted future satellite images complements using NWPs alone in making accurate solar energy predictions. Additionally, we propose that the task of forecasting future satellite images is pertinent to renewable energy generation forecasts and is a task which could be uniquely suited to be tackled with machine learning architectures used for AI weather forecasting.
How to cite: Fulton, J., Efremova, N., Simpson, N., Fenton, I., Corcoran, E., Robinson, J., Asthana, M., Yatsyshin, P., and Pedrazzini, N.: Using AI forecast of satellite imagery to improve solar generation forecasts, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18719, https://doi.org/10.5194/egusphere-egu25-18719, 2025.