A hybrid physical/statistical day-ahead direct PV forecasting engine
Photovoltaic (PV) penetration to the grid is growing rapidly, and more and more business models try to maximise their benefits in the context of smart grid services. Nevertheless, their volatile and intermittent behaviour remains a critical challenge for fully exploiting their potential while retaining overall grid stability and reliability, especially during planning day-ahead operation. Towards that direction, multiple PV generation forecasting algorithms and tools have been introduced by both research and industry, but without significant accuracy being achieved, especially for small scale PV installations. The presented approach combines a physical model to calculate the actual generation based on numerical weather forecast and PV plant technical specifications, and state-of-the-art machine learning algorithms to correct the error introduced by the limited accuracy of the online weather forecasting tools in the specific location of interested and using historical generation data. The proposed forecasting engine has been applied to a small-scale PV installation in Northern Greece. The engine’s performance is assessed via well-known metrics (mean absolute error, root mean square error), along with a newly designed metric that can better evaluate the examined challenge, the weighted relative squared error. Experimental results acquired demonstrate a very good performance rendering the proposed algorithm a very promising tool in the energy-related domain.