Re-modeling the solar diffuse fraction in Taiwan on basis of a typical-meteorological-year data

An objective approach to determine a dataset from which a correlation model of solar diffuse fraction is fitted and tested is proposed. A typical meteorological year (TMY) of global radiation data that is measured in Tainan, Taiwan over 10 years (2011–2020) is used to construct the training dataset and the remaining data (90% data base) is used as a test dataset. Two multiple-predictor and one single-predictor correlation models for the hourly diffuse fraction are developed by using more theoretically rigorous techniques for determining the breaking points for segmentation (excluding the modified Boland-Ridley-Lauret (BRL) model) and how many significant predictors required in each segmented interval for the regression model with piece-wise linear multiple-predictor correlation. The performance of each developed correlation model is superior to that of the existing same-type model using a part (one year or two years) of the same data base. The re-modeled piece-wise linear multiple-predictor correlation model has the best long-term performance of the three developed correlation models. The modified BRL model on basis of the TMY data is second. The re-modeled Liu-Jordan-type (single predictor) model allows real time prediction and has a simpler form than the other two models but prediction accuracy is inferior.