A Machine Learning-Enabled Clustering Approach for Large-scale Classification of Solar Data

Solar power generation is one of the renewable energy sources that has gained prominence in recent times. Power generation companies and researchers are interested in the behavior of solar facilities with respect to daily weather fluctuations over time. Researchers have often tried to cluster solar spatio-temporal data. Some of the proposed methods contain unnecessary ambiguity with no clear justifications of the applicability of their models. This study aims to provide an accurate cluster-based representation of solar Photovoltaic (PV) production profiles, using the solar data from the University of Louisiana at Lafayette's 1.1 Megawatt(MW) Photovoltaic Applied Research and Testing Lab (PART lab) over a two-year interval. The availability of these clusters provides an easier basis for testing and modelling the behavior and response of an existing solar plant in the event of an upgrade, or incorporation into a power generating facility and solar forecasting. In this study,a machine learning approach is used to generate a clustering algorithm, employing three different existing accuracy metrics. A efficient Dynamic Time-Warping (DTW) k-Means model was developed to cluster the PV data with reasonable accuracy. The novelty of this study provides an efficient temporal-adaptive multi-model approach of clustering solar PV data. These clusters can reduce the computational expense of solar-connected microgrid modelling and simulation studies at the design and operational stages.