Abstract

Solar radiation forecasting has become a critical information technology to facilitate the integration of PV and thermal solar power plants into the electricity grid of any country. Artificial neural network (ANN) modeling of time series is known as a useful and effective forecasting tool to achieve this task, due to its ability to find nonlinear relationships hidden inside historical data. Unfortunately, fast cloudiness transients add a stochastic signal to the solar radiation time series, thus diminishing the effectiveness of this methodology. In this work, ANNs are trained to provide 1-day-ahead forecasts of global solar radiation under various cloud regimes. Nine years of data measured under diverse climates at eight stations from the U.S. SURFRAD network are used. Training periods of less than two years are found too short and result in larger errors. Using a training period of eight years, the forecast accuracy is found to depend on cloud fraction (and thus location), with RMS errors ranging from 10% up to 45%.