Analysis of minute-scale variability for enhanced separation of direct and diffuse solar irradiance components using machine learning algorithms
Myeongchan Oh, Chang Ki Kim, Boyoung Kim, Chang-Yeol Yun, Jin Young Kim, Yong-Heack Kang, Hyun‐Goo Kim
Abstract
With the development of solar power generation, the importance of direct normal irradiance is being emphasized. Therefore, previous studies have examined numerous separation models for calculating direct and diffuse irradiance from global horizontal irradiance. These models have recently been applied to data in 1-min units. These have more information than the previous hourly data. This study proposes an advanced separation model for direct irradiance using machine learning (ML) techniques with 1-min time-series variabilities. Three ML models were proposed and trained using data from over 20 stations in temperate climate zone regions. The models were objectively compared and analyzed using untrained period data and validated using cross-validation. The results show that the proposed time-series ML models outperform the previous models in all the stations; moreover, the neural networks show the best results with a root mean squared error (RMSE) of 13%. This shows that an ML with enough data and variables is suitable for the separation model; especially, the influence of additional variables (i.e., variability) was found to be significant in reducing the RMSE to 30%. This study can be used as a base for separation models using ML techniques and variability.