Exploiting dimensionality reduction techniques for photovoltaic power forecasting

The availability of methodologies and tools to forecast the power produced by photovoltaic systems is of fundamental importance in many applications, such as the detection of anomalous events and the integration of these systems in the public electricity grid. In this paper we propose a novel approach to predict the produced power based on several weather variables. Similarly to previous works we model the power prediction task as a regression problem. However, in this paper, we rely on advanced machine learning algorithms such as Support Vector Machines empowered with nonlinear dimensionality reduction methods, in order to optimally exploit the correlation of the several weather variables and to filter out noisy variables. Our experiments, conducted on two different datasets corresponding to different solar panels, confirm the validity of the proposed method. With our approach the forecast and the measured values of power production have a good level of correlation, always superior to 0.9.