Sky Images for Short-Term Solar Irradiance Forecast: A Comparative Study of Linear Machine Learning Models

An accurate solar irradiance forecast is critical to the reliable operation of electrical grids with increasing integration of photovoltaic systems. This study compares short-term solar irradiance forecasts based on sky images using seven different linear machine learning algorithms. In the first step, several features are extracted from sky images, reconstructed, and next used as exogenous inputs to seven machine learning algorithms, i.e., linear regression, least absolute shrinkage and selection operator (Lasso) regression, ridge regression, Bayesian ridge (BR) regression, stochastic gradient descent (SGD), generalized linear model (GLM) regression, and random sample consensus (RANSAC). A representative dataset of three years of sky images with 1-minute resolution from 2014 to 2016 serves for comparison together with the clear sky indexes as inputs to forecast ground-level solar radiances for up to 30 minutes ahead. The results of the abovementioned algorithms are compared, where for 5 and 10 minutes ahead, Lasso has the highest accuracy with a root-mean-square error (RMSE) of 0.05 and 0.062 kW/m2, while for 15 to 30 minutes ahead, stochastic gradient descent provides the most accurate forecast with an RMSE of 0.067, 0.071, 0.074, and 0.076 kW/m2 for 15, 20, 25, and 30 minutes ahead horizons, respectively. For all the time horizons, Bayesian ridge is among the three most accurate models, and RANSAC has the highest error. The results show that ground-level solar irradiance can be forecasted with a relatively low average instantaneous error ranging from 0.05 to 0.1 kW/m2 depending on the model and forecasting horizon without imposing a too high execution time overhead, namely, less than 7 s. The accuracy of the forecast can be improved if combined with cloud detection algorithms. Overall, ridge, Bayesian ridge, and stochastic gradient descent provide more accurate forecasts for short-term horizons.