Delete search term


Quick navigation

Main navigation

Combining AI with Physical Models for Energy Loss Diagnostics in Solar Power Plants

Technical failures can lead to severe energy losses in solar power plants. One such failure occurs when the solar tracking system gets stuck at a certain position instead of rotating to optimally track the sun rays. In order to avoid such power losses, tracker faults must be detected and repaired in due time. Provided with enough historical data, an AI algorithm can be trained to analyze the power production data from a photovoltaic (PV) plant and automatically detect energy losses that can be attributed to tracker faults. One of the challenges for such an AI system is to do this under any weather condition, at any region of the world, but with no access to examples of historical tracker faults, which are typically rather rare. 

The challenge of having too little information about failures is not unique to PV plants. A novel approach to deal with this challenge is to bridge the gap of missing data using our knowledge of the system’s physics, in a so called “physics-informed AI” method.

In the project, a software module for the intelligent detection and diagnostics of energy losses in grid scale PV plants was developed. Understanding the tracker mechanism allowed the project team to develop a “fault generator” that uses data from a normal functioning operational solar plant, and “corrupts” it in the right way, such that it appears as if it came from faulty solar trackers. In a next step, the synthetic faulty data is fed together with the normal operational data into a deep learning neural network that is trained to distinguish between “faulty” and “healthy” solar strings.

The results of the classification demonstrate that the physics-informed AI improves the ability to detect tracker faults by 70% compared to an AI method which does not integrate the domain knowledge in it. Concretely, the hybrid algorithm managed to detect 392 out of 417 faults instead of detecting only 231 faults. This was achieved while keeping the false alarms close to zero.


Related papers:

  1. Goren Huber, Lilach and Notaristefano, Antonio, .
    Predictive Maintenance mit Physics-Informed-Deep-Learning : Anwendungsfall Photovoltaikanlagen. fmpro service. 2022(3), pp. 24-25. Available from:
  2. Zgraggen, Jannik; Guo, Yuyan; Notaristefano, Antonio; Goren Huber, Lilach, 2022. Physics informed deep learning for tracker fault detection in photovoltaic power plants.
    In: Kulkarni, Chetan; Saxena, Abhinav, eds., Proceedings of the Annual Conference of the PHM Society 2022.14th Annual Conference of the Prognostics and Health Management Society, Nashville, USA, 1-4 November 2022.Available from:
  3. Zgraggen, Jannik; Guo, Yuyan; Notaristefano, Antonio; Goren Huber, Lilach, 2023. Fully unsupervised fault detection in solar power plants using physics-informed deep learning.
    In: Brito, Mário P.; Aven, Terje; Baraldi, Piero; Čepin, Marko; Zio, Enrico, eds., Proceedings of the 33rd European Safety and Reliability Conference (ESREL 2023). 33rd European Safety and Reliability Conference (ESREL), Southampton, United Kingdom, 3-7 September 2023.Singapore:Research Publishing.pp. 1737-1745. 
    Available from:
  4. Goren Huber, Lilach; Palmé, ThomasArias Chao, Manuel, 2023.Physics-informed machine learning for predictive maintenance : applied use-cases
    In:2023 10th IEEE Swiss Conference on Data Science (SDS).10th IEEE Swiss Conference on Data Science (SDS), Zurich, Switzerland, 22-23 June 2023.IEEE.pp. 66-72. 
    Available from: