Solar Power Nowcasting in the Presence of Sahara dust: Can Deep Learning based on Satellite and Synthetic Production Data Recognize the Production-Offsets?

Fostering solar power as a sustainable, as fossil fuel free as possible, energy source demands accurate, location-optimized, and highly resolved forecasts of power production. For this study, we consider the issue of production offsets due to Sahara dust events in large areas of Central Europe as well as data coverage and inconsistency issues with evolving solar power sites. In the presented case study we investigate subhourly nowcasts using machine learning for (i) specific solar power plants in Central Europe (ii) the suitability of synthetically generated production using NWP grid points in the studied areas.  
Deep learning enables us to consider complex timeseries from historic data to model highly variable (spatial, temporal) diurnal and seasonal changes in the expected power production. In particular, we investigate how to exploit the spatio-temporal relationships in highly resolved forecasts by a sequence-to-sequence encoder-decoder inspired LSTM (long short-term memory) artificial neural network. We optimize the performance of our deep learning approach by tuning hyper-parameters, network weighting, and loss as well as addressing the input feature selection accordingly. Our preprocessing steps transform available data into a suitable representation for learning efficiently from a combination of multiple, very heterogeneous data sources with varying temporal availability and spatial resolution. For instance, we utilize 3D-fields from other weather prediction models, satellite data and remote sensing products, and observation time-series as well as their generated climatologies. A key objective is to properly process the differing temporal and spatial resolution while still generating nowcasts efficiently. To extend the historic training data set of complex models, we generate synthetic solar production data using machine learning and consider climatological driven data transformation. We investigate transfer learning as a further option in our deep learning setup.
Results obtained by the developed method generally yield high forecast-skills, where the best model setups are shown in our analysis. We compare the forecast results of up to 6 hours ahead obtained through this machine learning approach to available forecast methods, e.g., forecasts generated with python pvlib driven with AROME.