Use and adaption of ensemble learning methods to develop an irradiation nowcasting model with probabilistic output

Abstract

This master thesis proposes a method of creating probabilistic irradiance predictions for intra-hour situations by combining two ensemble prediction methods. These ensembles consist of historic measurements, which are chosen by similarity of the environmental situation and prediction results from machine-learning (ML) models, performing best with similar weather situations as the given data point. Hence the so accumulated data is processed by a natural-gradient-boost (NGB) model [13] to estimate a deterministic prediction, as well as an confidence interval. Since the location and width of such an interval is supposed to help estimating future irradiance values including uncertainty. The probabilistic predictions, generated by the NGB approach, generate superior results compared to the base ensembles by the continous ranked probability score (CRPS) as well as the mean-absolute-error (MAE). Mentionable is that the performance of the proposed method increases, by higher forecast horizon of 15 and 20 minutes with respect to the reference ensemble. This consists of an accumulation from an analog-ensemble [2] and an prediction-ensemble by a dynamic selection [11] process.