Development of a data-driven approach for the short-term prediction of solar power

Kurzfassung

The integration of solar energy system into the electricity mix requires a reliable forecast of its intermittency. The temporal variability of solar irradiance is caused by the dynamics of cloud cover over solar photovoltaic (PV) panels, which makes the PV output forecasting challenging. With the recent advancement of deep learning, machine learning techniques have shown a superior ability of performance in solar PV output forecasting and are then promising techniques to replace conventional physical approaches. In particular, computer vision-based approaches can extract and exploit the spatial and temporal information of the cloud movement. In addition, taking advantage of a growing number of datasets such as All Sky Imagers (ASI) and satellite observations, data-driven approaches are being developed in order to improve short-term and intra-hourly irradiance forecasts. In this study, we assess the feasibility of using a Convolutional Neural Network (CNN). A set of ASI, satellite images and past irradiance data are given at the input, while a series of future irradiance measurements (5 to 90 minutes ahead) are returned at the output. To assess the relative performance, the forecasting skill metric using root mean square error is used and the results are compared with the smart persistence model. The a-priori study shows that the CNN performs well as the prediction skill reaches about 15 ~ 19 % for predictions 5 to 30 minutes ahead.