Solar Irradiance Forecasting Under Cloudy Conditions Based on Statistical and Machine Learning Models

Accurate forecast of solar radiation, including total downwelling irradiance and its partitioning in the direct and diffuse directions, remains to be challenging under cloudy conditions due to the complex interactions between clouds and solar radiation. With the advancement of machine learning techniques, data-driven models allow us to investigate the inherent pattern of the relationship between clouds and radiation for better forecasting solar irradiances, especially in short forecast time-horizons (e.g., intra-day forecast). This study presents two statistical models, Holt-Winters model, and seasonal ARIMA model, as well as a deep learning model, Long-Short Term Memory (LSTM). Comparisons between the models with the persistence model are conducted using decade-long measurements of solar irradiances and cloud types (1998 to 2014) collected at the U. S. Department of Energy (DOE) Atmospheric Radiation Measurement (ARM) Southern Great Plains site to inspect the improvements in forecasting under eight cloudy conditons: shallow cumulus, other low clouds expect shallow cumulus clouds, deep convection, altocumulus, altostratus, cirrostratus/anvil, cirrus, and congestus. The interactions between clouds and radiation detected by the models may shed light on the seamless integration of data-driven and physics-based forecasting models, and thus improve numerical weather prediction models for solar forecasting.