Open Access

Oliver Dölle

• Energy systems worldwide are changing considerably with the ongoing expansion of renewable energy sources. Forecasts are imperative, to ensure efficient operation and trade with external entities in these complex decentralized energy systems. However, their inherent uncertainty can lead to forecasting errors and consequently to suboptimal operational plans and bidding behavior. One potential solution is to not only predict a single value, but to also estimate the existing uncertainty using probabilistic forecasts. In this context, probabilistic forecasts of generated photovoltaics (PV) power are particularly important, as its installed capacity is the fastest growing of all renewable energies. This thesis focuses on questions that still need to be addressed for a transition of probabilistic PV power forecasts to an applied industrial use in decentralized energy systems. To this end, the work of the author's corresponding publications is extended and supplemented, while establishingEnergy systems worldwide are changing considerably with the ongoing expansion of renewable energy sources. Forecasts are imperative, to ensure efficient operation and trade with external entities in these complex decentralized energy systems. However, their inherent uncertainty can lead to forecasting errors and consequently to suboptimal operational plans and bidding behavior. One potential solution is to not only predict a single value, but to also estimate the existing uncertainty using probabilistic forecasts. In this context, probabilistic forecasts of generated photovoltaics (PV) power are particularly important, as its installed capacity is the fastest growing of all renewable energies. This thesis focuses on questions that still need to be addressed for a transition of probabilistic PV power forecasts to an applied industrial use in decentralized energy systems. To this end, the work of the author's corresponding publications is extended and supplemented, while establishing an overarching context. Four approaches (e.g., mixture density network (MDN), generalized autoregressive model with conditional heteroscedasticity (GARCH)) that have yielded good results in solar irradiation forecasts or other forecasting fields are adopted and investigated in depth for PV power and compared to established methods. Additionally, a simulation with 24 different initializations and different amounts of training data is carried out in this thesis. Beforehand, there were no studies regarding the probabilistic prediction quality of PV power forecasts with limited amount of data, although this is indispensable for commissioning in practice. During the generation of the forecasts in this thesis, no manual intervention is applied, as this would also not be feasible in practice. Instead, several regularization methods are used. Furthermore, an automated time decomposition approach is developed for the autoregressive models with exogenous input (ARX), followed by a higher-level greedy search algorithm to determine the model order automatically. To represent the influence of possibly suboptimal model structures, extensions for modeling the epistemic uncertainty are implemented and analyzed for each approach. The simulations are conducted on the basis of PV power measurements from three sites in Central Europe spanning a period of around two years. The results indicate that even with seven days of training data, nearly all the methods show better forecast accuracies than the reference case of the complete history persistence ensemble. For all uncertainty representation forms the ARX-based probabilistic predictions are outperforming their respective neural network counterparts. Nevertheless, after six months of available days of training data, the behavior reverses and neural network approaches perform better on average. In general, the approaches with a continuous distribution have the best forecasting quality. Hence, the GARCH model in combination with the ARX model is recommended over the entire commissioning period, as it achieves excellent results both with a small (skill score: 31.4 %) and large (skill score: 34.3 %) amount of available training data in comparison. However, when provided with enough data, the MDN model surpasses the other methods in terms of overall forecasting accuracy with an improvement over the benchmark of 39.8 %.…