Data-Driven Model for PV Power Generation Patterns Extraction via Unsupervised Machine Learning Methods

Photovoltaic solar power generation varies from its nominal value over a wide range due to weather parameters intermittency and Sun path during a day. This intermittent behavior poses complications in the grid management systems while the solar power penetration rate grows incessantly. This paper concentrates on extracting photovoltaic power generation patterns to obtain a complete picture of dominant generation patterns at a daily time scale. In the proposed data-driven model, the high dimensional temporal features of the daily solar power output samples are transformed to the lower feature space through singular value decomposition, and then the K-means algorithm, the unsupervised machine learning technique, is applied to extract the five distinct dominant patterns. The results show that each pattern corresponds to different weather types such as sunny, mostly sunny, partially sunny, partially cloudy, and cloudy.