Entwicklung robuster Prognosen für ein Energiemanagementsystem anhand datenbasierter Modellierungsverfahren unter Berücksichtigung von Unsicherheiten

Abstract

The aim of this thesis is the development of robust and highly automated day ahead forecast methods for generation plants and storages. These form the basis for a transferable energy management system for agricultural companies. Data-based methods are formulated for both a deterministic and a probabilistic forecast. From a mathematical point of view, they only differ in the objective function of the underlying optimization problem. With a polynomial ansatz function deterministic and probabilistic day ahead forecasts for a photovoltaic plant, a wind plant and a battery storage are computed. Due to efficient optimization methods, both the training and the update of the models can be performed with significantly small computation times. In order to determine the achievable accuracy of the forecasts, detailed analyses are carried out on the basis of long-term data. This results in deterministic generation forecasts with an nRMSE of 12 % to 13 % and state of charge forecasts of a battery storage with an nRMSE of approximately 2 %. The comparison to other models in two benchmarks has shown that they lead to similar results. For the probabilistic models, other criteria are used to evaluate the quality of the forecast. The proposed quantile regression method leads to highly reliable forecasts of the generation plants, whereas the probabilistic battery storage forecast is less reliable due to the iterative evaluation of the dynamic model over the test horizon. Based on the general formulation of the forecast methods as optimization problems, the sensitivity analysis of the NLP-solver WORHP can be used to detect the most important input values for the generation forecast. In combination with the implemented model update, this results in a highly automated forecast. After a training duration of 20 days, these forecasts show similar results compared to the analysis of the long-term data. Again, this cannot be realized for the battery storage with the proposed iterative model evaluation of a static model. The forecasting models of the various components of the energy management system are finally combined in a simulation environment. Herein, the potential benefit of probabilistic generation forecasts for decisions in an energy management system is shown.