This study examines how different sources of information influence day-ahead solar photovoltaic (PV) energy generation forecasting at an industrial rooftop installation. Using nearly two years of hourly data from the ULMA Packaging factory in northern Spain, this study evaluates three practical input settings. These include production history, irradiance history derived from Solcast’s satellite-based observations and day-ahead irradiance forecasts provided by MeteoGalicia. A spectrum of models is evaluated, regularised linear baselines, feed-forward neural networks (FFNNs), and recurrent neural networks (RNNs), specifically the Gated Recurrent Unit (GRU) architecture. Each model is trained to predict the full 24-hour production profile in a single step with one hour resolution. The results show that no method dominates in every situation. Well-regularised linear models, especially Ridge, remain strong and surprisingly stable, performing best when fed with the forecast blocks. GRUs show mixed behaviour depending on the irradiance information setting and weather variability. FFNNs stay competitive throughout, generally falling within a small margin of the best models. Weather conditions play a major role: errors are lowest and most uniform on sunny days, moderate on partly cloudy days, and highest under cloudy skies, where the quality of irradiance information becomes the limiting factor. Overall, the study highlights that the choice of input information can matter as much as the model architecture itself. Distinguishing past information from deployable forecasts offers a clearer picture of what can realistically be achieved in operational Energy Management Systems (EMS).
Pour et al. (Mon,) studied this question.