Robust ML-nowcasting of solar irradiance from satellite derived features

With the increasing shift to renewable energy sources, their predictability is becoming more of a concern. While some renewables follow more stable power output patterns, solar irradiance and thereby photovoltaic power production can experience significant shifts in shorter timeframes. The great potential of solar irradiance as a power source should although not remain underused, as it globally provides orders of magnitudes more energy to the earth than currently or foreseeable required. Solutions lie in forecasts for different timescales, with the short-term and nowcasting domains being able to provide the most accurate insights to volatility introduced by atmospheric phenomena. These forecasts are not only relevant to determine the expected economic impact but also to maintain an equilibrium in electrical grids with the goal of minimizing the waste of potential power production from solar irradiance.

Large-grid-nowcasts of solar irradiance can substitute forecasting of solar power potential for individual sites, which are typically derived from the sites’ measurements. With models using satellite data instead, forecasts for large-areas are available, which are useful to approximate the solar intensity for a range of increasingly spatially distributed photovoltaic power stations. Satellite data is in contrast to ground-based data sources or NWP model estimates less reliant on the proper workings of a wide range of externalities and readily available in near-real-time.

Based on a study of multiple convolutional-recurrent neural network architectures, deriving nowcasts from a single solar irradiance satellite-data-product, further research is undertaken to determine the limitations and benefits of single-irradiance-feature nowcasts and counteract potential detriments.

A known issue with this kind of pipeline lies in its main benefit: The easy near-real-time access to a single dynamic feature can bring the whole model to a halt if an issue occurs with this single externality. To gather insights and provide a practical solution to this problem a further study on model robustness to missing data is undertaken, leading to the technique Timestep-Dropout. Via probabilistic removal of irradiance frames during training, neural networks can learn to expect missing frames in inference and still derive forecasts from the remaining valid frames.

Possible benefits in forecast accuracy through the introduction of further features from satellite data or data from other sources may outweigh the additional burden of requirements, and provide overall improvements. To estimate these effects a comparison to the multi-irradiance-feature trained model will provide insights into a reasonable balance of irradiance-feature requirements.