Improving sky image-based short-term solar irradiance forecasting through the integration of meteorological data

Fluctuations in solar irradiance (SI) cause instability in power generation, challenging grids with high photovoltaic (PV) penetration and impacting power quality. Accurate short-term SI forecasting is crucial for reliable solar energy integration, enabling grid operators to optimise dispatch, manage reserves, and schedule energy storage. It also supports compliance with electricity markets requiring hourly or sub-hourly scheduling. This study presents a hybrid deep learning model combining convolutional neural networks (CNN) with long short-term memory (LSTM) networks to forecast SI 60 min ahead. The model uses images from the Karlsruhe low-cost sky imager (KALiSI) alongside meteorological data from a local weather station. Correlation analysis identifies key meteorological variables associated with global horizontal irradiance, which are integrated with sky images to improve forecasting accuracy. Experimental results show that including relative humidity (RH) and wind direction (WD) yields the lowest average root mean square (RMS) error of 119 W/m² and mean absolute error (MA) error of 82 W/m². This outperforms image-only and persistence models by 3 % and 42 % in RMS error (9 % and 50 % in MA error), respectively. Adding more meteorological variables does not enhance accuracy and may sometimes degrade performance, emphasising the importance of informed feature selection. At longer horizons (up to 120 min), all models show increased errors, with persistence performing worst (normalised RMS error 60 %, normalised MA error 47 %). In contrast, the image-only and RH+WD models maintain more stable errors (normalised RMS error 34–35 %, normalised MA error < 27 %). These results pave the way for improved forecasting for solar power production. • Solar irradiance (SI) fluctuations can hinder stability in photovoltaic power grids. • A CNN-LSTM model forecasts SI 60 min ahead using sky images. • Including relative humidity and wind direction improves forecast accuracy. • Extra meteorological inputs can reduce accuracy of sky image-based forecast models. • Strategic feature selection is essential for deep learning models in SI forecasting.