Forecasting rare but impactful events in renewable energy generation - condition classification for optimal expert model training and model selection in PV forecasting

PV generation is affected strongly by the short-term fluctuations in meteorological conditions - from clear sky to cloudy, which would be considered normal conditions, to more rare events like snow fall, freezing rain or dust that covers the modules. Rare events are hard to forecast reliably, they are usually not well represented in the training data and cause imbalances in training and prediction. However, they can have huge impacts on the energy system as for example snowfall usually covers many PV plants simultaneously over extended areas. Such events, if not foreseen in time, require balancing action of network operators and thereby cause large costs and possibly strain on the energy infrastructure.

State of the art PV forecasting models are overwhelmingly being trained on datasets without accounting for changing conditions and rare situations. To improve the prediction of such events we present a new method in the form of a data labeler and classifier for snow conditions in PV forecasting based only on meteorological and historical PV generation data to allow for a classification of the expected forecasting conditions over a time horizon of the next few hours up to days. With the classification performed, the best suited model trained for the expected condition can be employed to yield the most reliable PV forecast.

The method is site-specifically trained with historical PV generation data of the site, but no other metadata, module specifications, satellite or visual data are required. The data is combined with historical weather measurements (like irradiance, temperature and precipitation) from a close-by meteorological station. By classifying the conditions in the training dataset with the method, rare conditions are identified and labelled. The labels do not require exact validation, a high likelihood is sufficient. Expert models for those conditions can then be trained in a supervised setting. These are exposed to a training dataset that has dense samples of the selected rare condition and can include augmented samples of the condition. Thereby, a range of specialized forecasting models is created and benchmarked against each other to ensure selection of the best performing models for forecasting in case of a forecast rare condition.

Preliminary results from Austrian PV systems indicate a high accuracy of 99.6% and true positive rate of 96% for the labelling method with a false positive rate of only 0.05% on a test dataset. An LSTM neural network-based classifier to forecast conditions 24 hours ahead shows similar performance metrics and an LSTM regressor expert model achieved only 30% of the PV forecasting error of a similar non-expert model. Both classifier and expert regressor were trained on the labelled and condition enriched dataset.

The work was funded by the Austrian Climate and Energy Fund and carried out under the program "Energieforschung 2022".