Detection of shading for short-term power forecasting of photovoltaic systems using machine learning techniques

This paper presents a mahcine learning based solar power forecast methode that can take into
account shading related fluctuations. The generated PV power is difficult to predict because there are various
fluctuations. Such fluctuations can be weather related when a cloud passes over the array. But they can also
occur due to shading caused by stationary obstacles, and this paper addresses this form of shading. In this work
an approach is presented that improves the forecast under such fluctuations caused by shading. A correction of
the prediction could successfully reduce error due to shading. The evaluation of the model is based on five sets of
recorded shading data, where shading resulted from intentionally placed structures. The correction uses internal
inverter data and irradiance values of the previous day to perform the correction and was able to reduce the
RMSE of four 10 kWp systems with different orientation and tilt angle under shading and thus improve the
prediction accuracy by up to 40%. The model can detect how intense the shading is and correct the forecast by
itself.