Machine learning approach to reconstruct fog history for solar parks

Fog or low stratus clouds can reduce solar power production significantly. A not forecasted fog event can lead to huge power misses, even at country scale and thus lead to high imbalance prices in energy trading. This presents a financial risk to the solar park owner, who bears the balancing responsibility in the European power markets. Reducing this risk makes solar investments financially more interesting and thus can help for the achievement of a climate neutral Europe. More accurate solar forecasts can help to achieve this goal, and for this an accurate historical data of fog occurrence at the park level is important. It is not straightforward to get this information for several reasons: Fog can be a very local phenomenon, and observations only rarely exist close to solar parks. Reanalysis data such as ERA5 do not have visibility information and numerical weather prediction models (NWP) often poorly predict fog, in part because their grid is too large to model fog formation. &#160;In this study, we will present a machine learning model to reconstruct fog history at solar park level. To build our model we use solar parks, which have weather observation in close vicinity. The weather observation will give us information about the true fog events and thus the target data, i.e. the data we want to predict. We use visibility and low cloud measurements from observation stations from the German weather operator DWD, Deutscher Wetterdienst. Fog was defined to include fog and mist (visibility < 5km) and to low clouds, which are lower than 2 km. To train the model, we will use reanalysis data from ECMWF (ERA5) and solar production. Part of the work was to find the best features to represent fog formation. We used data such as cloud cover, temperature (at surface and gradients) and humidity. Then we trained a classifier to predict the fog history at park level. This model showed good improvements of fog detection compared to NWP fog forecast. Our model achieved an accuracy of 94% and Brier score of 0.05, while the NWP fog forecast of the closest grid point to the solar park shows respective values of 86% and 0.17. This shows that our model was able to produce accurate historical fog data at the solar park level.