Very short-term solar irradiance forecasting based on open-source low-cost sky imager and hybrid deep-learning techniques

Solar irradiance (SI) forecasting is vital for reliable photovoltaic (PV) operation. This is especially true for regions like Africa where many SI forecasting approaches rely on scarce historical data and the inherent instabilities of electric grids are further compounded by SI variability. Accurate solar forecasting is essential for improving grid management, enabling operators to balance supply and demand and enhance stability. Ground-based sky imaging is a promising technique for SI forecasting that do not require extensive historical data. However, commercial sky imagers are expensive and offer limited flexibility. This paper introduces the Karlsruhe low-cost all-sky imager (KALiSI), made from off-the-shelf components that captures high-resolution images and can be assembled for less than €600. The KALiSI was installed in Karlsruhe, Germany, to collect images to train a convolution neural network-long short-term memory (CNN-LSTM) model for 15 min-ahead forecasting of global horizontal irradiance (GHI). The root mean squared (RMS) error of the model ranges from 19–206 W/m2, compared to 33–257 W/m2 for persistence, while mean absolute (MA) errors range from 15–144 W/m2 for CNN-LSTM and 30–159 W/m2 for persistence. The model’s performance using KALiSI’s images was compared with a commercial sky imager at the same location across various forecast horizons. The KALiSI showed normalised RMS error and MA error values of 6 % and 7 % higher, respectively, with some discrepancies noted on clear days. These results show the KALiSI’s suitability for very short-term forecasting and its open-source design offers a low-cost solution for developing countries.