Inscrição na biblioteca: Guest

ISSN Online: 2377-424X

ISBN Print: 978-1-56700-421-2

International Heat Transfer Conference 15
August, 10-15, 2014, Kyoto, Japan

Forecasting Methods for Direct Normal Irradiance at the Ground Level

Get access (open in a dialog) DOI: 10.1615/IHTC15.sol.009309
pages 7633-7647

Resumo

Hybrid models combining stochastic learning methods with cloud image processing are used to develop shortterm forecasting systems for Direct Normal Irradiance (DNI). The fully automated forecasting system is used to predict ground irradiance at time horizons that are relevant to Real-Time Dispatch (RTD) operations of concentrated solar power plants (5 and 10 minutes ahead of forecast issue time). We discuss the process of constructing hybrid models that perform well for both high and low variability seasons of solar irradiance. Six months of diurnal imaging data taken every 30 seconds for a Central California location with a sky imager are used to build the deterministic sector model that forms the basic image translation component of the method. The overall methodology employs sky image processing, deterministic and Artificial Neural Network (ANN) forecasting models, a genetic algorithm (GA) for optimal ANN topology and input selection, and a number of different training methods, including a newly developed randomized training and validation method (RTM). Forecast performance for each season is evaluated in terms of standard error metrics and a suitable forecasting skill (s). The smart models show signicant improvement over the deterministic and persistence models. Based on the models with higher forecasting skill for each season, we build a season-independent hybrid forecast model for DNI that can be easily adapted to power output from Concentrate Solar Power (CSP) and Concentrated Photo-Voltaic (CPV) plants. The hybrid model achieves 20% forecasting skill improvement over diurnally-corrected persistence for both time horizons of 5 and 10 minutes.