Optimal selection of training datasets for solar nowcasting models

Most solar nowcasting approaches use regression analysis to fit a model to historical solar irradiance data which predicts solar irradiance in near-real time (e.g. intra-hourly). A priori, any amount of historical data can be used as training material, e.g. months to years. However, evidence suggests the hypothesis that fitting a solar nowcasting model to a data partition that presents some form of solar irradiance cohesion (e.g. clear vs. cloudy sky)can help select the appropriate forecasting technique (e.g. linear vs. non-linear) to achieve increased accuracy. In this paper, we present a systematic study to test this hypothesis by (a) developing different methods to select solar data partitions to train nowcasting models, (b) assessing the relative performance of each method by evaluating the same forecasting algorithms with the different data partitions, (c) developing an optimal technique to identify and match partitions in solar time series data with forecasting algorithm