Short-Term Photovoltaic Power Forecasting Using Sky Images

The increasing deployment of photovoltaic systems is driving the need for accurate methods to forecast electricity generation over short time horizons. In particular, the availability of sky images makes it possible to incorporate information on cloud cover and its evolution into forecasting models.

This thesis focuses on the development and evaluation of techniques for short-term photovoltaic power forecasting (up to approximately one hour ahead) using sky images. The dataset includes sky images associated with photovoltaic power generation measurements and other variables relevant to the forecasting task.

The work may involve the development of models based on Machine Learning and Deep Learning techniques for image analysis, potentially also accounting for the temporal evolution of sky conditions. Their performance will be compared with that of more traditional forecasting approaches based on time series and meteorological variables, as well as with hybrid approaches combining image-derived information with numerical data.

An additional objective of the thesis will be to investigate the relationship between the amount of available data and model complexity, assessing under which conditions Deep Learning techniques provide an advantage over simpler approaches. The study may therefore include experiments with progressively smaller training datasets to characterize the accuracy, data requirements, and computational cost of the different methodologies.