Identifying representative days of solar irradiance and wind speed in Brazil using machine learning techniques

Abstract

The investment levels in electricity production capacity from variable Renewable Energy Sources have substantially grown in Brazil over the last decades, following the worldwide-seeking-goal of a carbon-neutral economy and the country’s incentives in diversifying its generation mix. From a long-term perspective, the current non-storable capability of renewable energy sources requires an adequate uncertainty characterization over the years. In this context, the main objective of this work is to provide a thorough descriptive analytics of the time-linked hourly-based daily dynamics of wind speed and solar irradiance in the main resourceful regions of Brazil. Leveraging on unsupervised Machine Learning methods, we focus on identifying similar days over the years, Representative Days, that can depict the fundamental underlying behaviour of each source. The analysis is based on a historical dataset of different sites with the highest potential and installed capacity of each source spread over the country: three in the Northeast and one in the South Regions, for wind speed; and three in the Northeast and one in the Southeast Regions, for solar irradiance. We use two Partitioning Methods (K-Means and K-Medoids), the Hierarchical Ward’s Method, and a Model-Based Method (Self-Organizing Maps). We identified that wind speed and solar irradiance can be effectively represented, respectively, by only two representative days, and two or three days, depending on the region and method (segments data with respect to the intensity of each source). Analysis with higher Representative Days highlighted important hidden patterns such as different wind speed modulations and solar irradiance peak-hours along the days.