Evaluation of Time-Series, Regression and Neural Network Models for Solar Forecasting: Part I: One-Hour Horizon

The need to forecast solar irradiation at a specific location over short-time horizons has acquired immense importance. In this paper, we report on analyses results involving statistical and machine learning techniques to predict hourly horizontal solar irradiation at one-hour ahead horizon using data sets from three different cities in the U.S. with different climatic conditions. A simple forecast approach that assumes consecutive days are identical serves as a baseline model against which to compare competing forecast alternatives. One approach is to use seasonal ARIMA models. Surprisingly, such models are found to be poorer than the simple forecast. To account for seasonal variability and capture short-term fluctuations, cloud cover is an obvious variable to consider. Monthly models with cloud cover as regressor were found to outperform the simple forecast model. More sophisticated lagged moving average (LMX) models were also evaluated, and one of the variants, LMX2, identified at monthly time scales, proved to be the best choice. Finally, the LMX2 model is compared against artificial neural network (ANN) models and the latter proved to be more accurate. The companion paper will present algorithms and results of how such models can be used for 4-hr rolling horizon and 24-hr ahead forecasting.