Solar irradiance forecasting in the tropics using numerical weather prediction and statistical learning. (1st March 2018)

• Record Type: Journal Article
• Authors: Verbois, Hadrien; Huva, Robert; Rusydi, Andrivo; Walsh, Wilfred
• Is Part Of: Solar energy. Volume 162(2018)
• Page Start: 265
• Page End: 277
• Publication Date: 2018-03-01
• Subjects: Solar energy -- Periodicals; Solar engines -- Periodicals; 621.47
• DOI: 10.1016/j.solener.2018.01.007
• Languages: English
• ISSNs: 0038-092X

Abstract

Highlights: Three configurations of WRF are used to forecast solar irradiance in Singapore. In comparison with interpolated GFS forecasts, RMSE is decreased by 3.4–10.1%. A new multivariate post-processing procedure is applied to the NWP models. It significantly reduces the RMSE and outperforms the univariate approach. Our best model gives significantly better results than tested benchmark models. Abstract: Increasing penetration of distributed renewable power means that reliable generation forecasts are required for grid operation. The present work aims at combining state of the art implementations of the Weather Research and Forecasting (WRF) model with multivariate statistical learning techniques to provide the most accurate forecasts of day-ahead hourly irradiance in Singapore. Three implementations of WRF-including WRF-solar-were used to produce three years of hourly day-ahead irradiance forecasts. Their performances were compared with that of the Global Forecasting System (GFS), which was interpolated to provide hourly forecasts. A multivariate post-processing procedure combining Principal Component Analysis (PCA) and stepwise variable selection was developed and applied to the four models. A smart persistence model and a climatological forecast were also implemented and served as benchmarks. The skill of the various models were evaluated using several metrics and statistical tests. We found that WRF-solar combined with our proposed statistical learning method outperformed smart persistence, a climatological forecast and GFS for day-ahead forecasts of irradiance. In particular, our model was shown to have a Root Mean Square Error (RMSE) 23% lower than smart persistence.