Drawing on historical data that is already available as a result of operating a solar power plant as well as specialist knowledge about the facility, the ZHAW researchers put together a neural network for intelligent fault diagnosis that can be integrated into the existing Nispera platform. In order for the neural network to perform reliably and intelligently, it needs a lot of data relating to faults, such as tracker errors. However, such events appear very rarely in the inventory data. With the help of the knowledge of the physical laws of the plant and the specialist knowledge of the engineers working on it, the team was able to corrupt the real data and thus create artificially “falsified data” containing realistic fault patterns with which the algorithms could be trained.

The researchers were thus able to develop a reliable software module without the need for elaborately generated training data by applying the approach fittingly known as physics-informed AI. This type of predictive maintenance also represents a novelty in terms of the maintenance of PV power plants. “By combining valuable specialist knowledge and a data-driven deep learning model, the module is able to identify faults in the system at an early stage and more accurately than was possible with previous approaches,” says Goren Huber in summarising the program’s advantages. “The ability to detect tracker errors has been improved by 70 percent relative to conventional AI models, meaning that the software module can contribute to reducing energy losses and ensuring the more cost-efficient maintenance of solar power plants,” adds the ZHAW researcher.