Solar energy is becoming one of the world's most important sources of electricity. In Lithuania, the number of solar power plants is also growing rapidly, and the energy they produce is playing an ever more significant role in the country's energy system. At the same time, however, one of the key challenges is becoming more pronounced: solar power generation remains highly dependent on weather conditions.
Even a small cloud briefly covering the sun can reduce a solar module's power output by tens or even hundreds of watts within seconds. When a large number of solar power plants are connected to the grid, such sudden fluctuations become a challenge for electricity grid operators, who must ensure a constant balance between generation and consumption.
A team of researchers at Kaunas University of Technology (KTU), led by Professor Rytis Maskeliūnas, has proposed a new solution to this problem: an artificial intelligence-based system called ShadowSense. The system is learning by observing the sky while simultaneously recording changes in a solar module's power output.
Short-term forecasts covering the next few minutes can help manage energy storage systems more efficiently, plan the operation of solar power plants more accurately and reduce the need for reserve capacity. The earlier the system understands that clouds are about to reduce output, the more time there is to prepare for the change.
Installed and Tested on a Roof of a Residential Building
"Most image-based artificial intelligence methods require large amounts of manually labelled data. This means that clouds, shadows or other objects in images have to be labelled manually. This process is lengthy, expensive and difficult to adapt to different locations or weather conditions," says Maskeliūnas.
ShadowSense works differently: the system analyses images of the sky, observes cloud movement and links this information to changes in the solar module's power output recorded at the same time.
"Each sudden drop in power becomes a kind of clue for the system, helping it understand which changes in cloud cover may have caused it," explains KTU professor Maskeliūnas.
In this way, the artificial intelligence learns not from manually labelled images of clouds, but from the real relationship between the appearance of the sky and the operation of the solar module.
"It was important for us to develop a system that could adapt to a specific environment. Every solar power plant has its own position, module angle, surroundings and local weather conditions. This means that a single universally prepared model may not always be accurate enough," says the KTU researcher.
The system first analyses images of the sky and the direction of cloud movement. It then assesses the position of the sun and calculates how cloud shadows may affect the solar module. This information is linked to previous changes in power output, enabling the model to predict how output will change over the next minute or several minutes.
One of the most interesting aspects of the study is that the experimental system was installed not in a laboratory, but in a real residential setting in Kaunas. A wide-angle camera was mounted on the roof of a building, regularly capturing images of the sky, while a solar module installed in the courtyard supplied power to the measurement system and the artificial intelligence computer.
Over 92 days of observations in Kaunas district, more than 122,000 synchronised observations were collected. Each consisted of a sequence of sky images and solar module power data recorded at the same time.
"We wanted to test whether such a system could operate not under ideal laboratory conditions, but in an everyday environment – the kind of setting in which solar power plants are actually used. This is important because clouds, light levels and weather conditions in the real world are constantly changing," says Maskeliūnas.
Operates in Real Time and Detects Sudden Changes in Power Output
The results of the study showed that ShadowSense can predict short-term changes in solar module power output more accurately than conventional methods. The system reduced the average forecasting error by almost a third and detected more than 92 per cent of sudden power changes associated with cloud shadows.
"For electricity grid management, it is important not only to know that a solar module will generate less energy, but also to understand as early as possible when this will happen. Even a short warning of several tens of seconds can be significant when generation, storage and consumption need to be balanced," notes the KTU researcher.
The system's efficiency is equally important. A single forecast took around 66 milliseconds to calculate, while energy consumption amounted to about 0.52 J per forecast. This means that the system can operate in real time on a low-power computer.
Such a solution could be relevant for decentralised solar power plants, remote sites or systems where powerful servers or a continuous internet connection are not available.
According to the KTU researcher, technologies of this kind could in future be important not only for individual solar power plants, but also for the broader energy system. As more renewable energy sources are connected to the grid, the ability to accurately forecast their fluctuations in real time becomes increasingly important.
"Future energy systems will not only have to generate clean energy, but also be able to respond to their surroundings. Technologies such as ShadowSense allow solar power plants to become smarter – they not only generate energy, but also learn to understand what is happening around them," says Maskeliūnas.
The article "ShadowSense: Edge Optimized Self-Supervised Learning for Dynamic Shadow Mapping of Solar Panels" can be found here .
