Spanish scientists developed an innovative system for solar radiation forecasting based on deep structured learning and using the sky images and numerical meteorological data. The new method allows for assessing the amount of solar energy available in the next 10–60 minutes with high level of accuracy. It is especially important for stable and efficient operation of solar power plants.
Solar radiation forecasting is a critically important task, because solar power plants are directly dependent on radiation intensity. The key problem is high versatility of the solar stream depending on the time of day, season and a series of meteorological factors. Even if a cloud appears for a short time, this may immediately decrease the generation and result in cuts in energy supply. Until now, the forecast was based on physical-mathematical simulation or on analyzing numerical meteorological data, but such forecasts did not take into account visual parameters of the atmosphere and required a big number of sensors. This problem was partially resolved with the help of neural networks, however, the majority of the existing models were based on just one type of input data — either images, or numerical parameters, which limited the accuracy of the forecasts.
The researchers from Almeria University and CIEMAT research center operating on the basis of the Solar Energy Platform (the biggest in Europe) proposed a multimodal approach. Their model simultaneously processes the sky images and numerical parameters, such as dates and theoretical non-terrestrial radiation, which allows for combining short-term visual signals, such as appearing clouds, with seasonal and geometrical parameters including the position of the Sun. The model was trained based on the data collected at the pilot CESA-I unit — one of the leading solar platforms in Europe. Between April 2022 and September 2023, more than 100,000 photos of the sky were received with the help of wide-angle cameras, and each photo was synchronized with solar radiation measurements and other meteorological parameters. Hel-IoT, an especially developed interface was used for preliminary processing and marking the images. It allowed for efficient formation of training samples for neural network models.
To avoid misrepresentation, the researchers classified the days based on the cloud cover degree dividing them into three groups: clear, cloudy and alternately cloudy. For that they used Clarity Index (CI), reflecting the ratio of the actual and theoretically possible solar radiation. Seasonal variations were also taken into account; however, it became clear that the model itself was capable of accounting for them, subject to receiving such data as dates and theoretical non-terrestrial radiation. The pre-trained ultra-precise neural networks (EfficientNet, ResNet and Xception) were used to process the images by extracting visual signs of cloudiness. The obtained characteristics were then passed to the recurrent LSTM network capable of taking into account the cloud cover dynamics. Simultaneously, numerical time series were analyzed with the help of NARX model. The entire system was integrated into a single hybrid architecture capable of forecasting the solar radiation level in 10, 30 and 60 minute horizons.
As a result, the multimodal approach demonstrated high level of accuracy: the 10-minutes-ahead forecast had an error of only 2.5–5 %, and one-hour-ahead forecast — about 10 % on average. The advantage was especially noticeable in alternate cloudy conditions, when the multimodal model was significantly better than the solutions based only on meteorological data or only on images.
Further on, the researchers plan to extend the functionality of their model by way of integrating satellite images, using the computer vision methods for automated calculation of clouds and generative neural networks — for creating realistic images of clouded sky
