Short-term forecasting of solar irradiance

Five statistical models for nowcasting solar irradiance are evaluated from different perspectives. The first four models are purely statistical ones: random walk, moving average, exponential smoothing and autoregressive integrated moving average. These models can be considered as benchmarks of different levels of complexity. The fifth model is a version of the two-state model, an applications suite for nowcasting solar irradiance developed by our team. The two-state model connects in an innovative manner an empirical estimator for clear-sky solar irradiance with a statistical predictor for the sunshine number, a binary indicator stating whether the sun shines or not. On the basis of different error metrics, the models’ performances are analyzed from four perspectives: forecast accuracy, forecast precision, data series granularity and variability in data series. The study is conducted with high-quality radiometric data measured at a high frequency of four samples per minute on the Solar Platform of the West University of Timisoara, Romania. No model is ranked as the best, but the peculiarities that cause a model to perform better than others are discussed. By processing information about the atmospheric transmittance, the two-state model proves a slight advance in the forecast accuracy and a notable performance in the forecast precision.