Smart Cloud Detection System for Intra-hour Solar Irradiance Forecasts (Invited Presentation)

This study proposes a Smart Adaptive Cloud Identification (SACI) method for sky imagery and solar irradiance forecast. The SACI method first uses a smart adaptive thresholding set (SAT) that takes into account the time series of global horizontal irradiance (GHI) and a frequency ratio map of the input image to classify sky images into three categories: clear, overcast, and partly cloudy. An optimal cloud detection scheme is then applied to each image category. This overall classification is optimized through supervised learning to create a library of features. The supervised learning outperforms all reference cloud detection methods used for our many solar observatories in 6 different microclimates with an overall accuracy higher than 90%. After that, a grid-cloud-fraction method is used to quantify the cloud cover information, and the numerical cloud cover information is used as inputs to artificial neural networks (ANNs) to forecast one minute-average GHI values. Three months of one-minute imaging data and half-minute irradiance measurements are used for model estimation and validation. The performance of this stochastic learning forecasting model is assessed in terms of common error statistics (mean bias and root mean square error), but also in terms of forecasting skill over persistence. This automated image-based ANN model achieves forecasting skills of 14%, 18%, and 19% over persistence forecasts for 5, 10 and 15-minute horizons.