Artificial intelligence for solar power plants
aiDrones and machine learning allow identification of modul defects within seconds.
12/2020 - HI-ERN Scientists use artifical intelligence to find module defects within seconds
Solar systems are on the rise worldwide. Ever larger shares of the power demand are covered by solar cells, and ever larger solar fields are being built for this. The largest of these is in the province of Qinghai in China. It has a capacity of 2.2GW, several million PV modules are installed there. But there are also many power plants with hundreds of thousands of modules, for example in Senftenberg or Templin. The inspection of such large-scale plants requires innovative concepts with a high degree of automation. Even if the inspection of a single module takes only a second, it takes days for 100,000 modules. Actually, it should be even faster, but even at one second per module, evaluations can no longer be done manually.
Up to now, however, plant inspections are still largely manual work. With image-based evaluation procedures, measurements are often carried out with a drone, which is steered through the module field by a pilot. The images obtained must then be post-processed before specialists evaluate them to identify module, string or cell defects. Such a procedure is time-consuming and therefore expensive. It is thus not surprising that solar field operators often turn away when drones are mentioned.
The COSIMA project is intended to show that there is another way. In this project, which is financially supported by the Federal Ministry of Economics and Energy BMWi, the Helmholtz Institute Erlangen-Nuremberg for Renewable Energy, or HI ERN for short, is working with the Nuremberg energy supplier N-ERGIE AG, the electronics company Automatic Research GmbH, the camera manufacturer IRCAM GmbH, Rauschert Hennigsdorf-Pressig GmbH, DHG Engineering GmbH and the Technical University of Nuremberg as well as the Allianz Zentrum für Technik AZT (Allianz Risk Consulting GmbH) on a process that relies on autonomous drone technology and artificial intelligence.
The goal of the project is a completely automatic and fast system inspection. The vision - an autonomous drone flies a pre-programmed course automatically and records a video of the photovoltaic system with a special camera. This camera detects light emitted by the system and generates images containing detailed information about defects in solar cells and modules. The images of the individual modules are automatically processed and evaluated by a pre-trained neural network to detect different error patterns. For conspicuous modules, a power simulation is performed to quantify how much they will reduce or decrease the power yield and financial return. The result then leads to a recommendation for action, for example: the module must be replaced, or shading should be removed.
For research purposes, HI-ERN has been using parts of this method for some time and has already analyzed tens of thousands of modules and found numerous defects. The biggest challenge for automation is to create suitable sample data sets for different types of defects. For a neural network to be able to detect a certain error pattern with high reliability, data from hundreds of examples of this defect must be available. The corresponding images can either be generated with special measurements in the laboratory or have to be marked manually in existing data sets. This procedure is cumbersome, but once an error type is characterized, the corresponding defect can be identified extremely quickly.
In the laboratory, for example, this is already being used to analyze crack formation. For this purpose, luminescence images of solar modules with cracks and fractures are recorded. These images take advantage of the fact that solar cells and LEDs are very similar in principle, only the mode of operation is reversed. If an electrical voltage is applied to a solar cell, it emits light. Due to the material used, silicon, this light is not visible to the naked eye, but an infrared camera can detect it. The images generated in this way are very suitable for detecting defects in solar cells, especially cracks and cell breakage. 744 luminescence images of modules with cell cracks and breaks were taken at HI ERN and used to train a neural network. This network was subsequently not only able to automatically detect broken cells in a module, but was also able to predict the associated loss of power. But that's not all - solar modules in a PV system do not produce electricity in isolation, but in combination with many other modules. Because the modules are interconnected, the operating point of a defective module in the network is different than if it were operated alone. A defective module in isolation can still have over 90% of its original power output, but only contribute about 50% when connected together. Also the contribution to the power in the network could be predicted well with a neural network. Within the framework of COSIMA, this approach is to be transferred to thermographic images taken by a drone and automated.
A fully automated PV system inspection is still a vision, but we are getting closer to it every day. At the beginning of October 2020, N-ERGIE AG demonstrated how it intends to implement this vision. At a press event, the camera drone, which was co-developed in the project, was presented with a test flight over one of the in-house photovoltaic systems. Autonomous operation is to be demonstrated by the end of 2021.
Dr. Ian Marius Peters, Dr. Claudia Buerhop-Lutz, Lukas Bommes, Dr. Jens Hauch
Helmholtz Institute Erlangen-Nürnberg for Renewable Energy (HI ERN)
Magazin "Sonnenenergie.de" reports about "Artificial intelligence for solar power plants" in January 2021