Refine
Document Type
Has Fulltext
- no (2)
Is part of the Bibliography
- no (2)
Keywords
- Hindcasting (1)
- Machine Learning (1)
- Photovoltaic (1)
- Photovoltaic Systems (1)
- Solar Irradiance (1)
Institute
This work introduces an innovative approach to calculate high-accuracy solar irradiance data for effective asset management of photovoltaic plants using Machine Learning. Ground-based pyranometers are expensive and seldom maintained, while weather service providers face limitations in spatial and temporal accuracy. A novel irradiance data model is introduced, that combines satellite weather information with data from PV plants to reconstruct historical irradiance levels with high accuracy. Our method uses existing PV arrays as "virtual sensors" to capture the local operating conditions, specifically the local irradiance incident on the array. The model was developed and validated using data from 43 medium to large-scale PV plants and two high-precision irradiance sensors. Results show superior performance compared to satellite weather data. With a root mean square deviation of 71 W/m² for global horizontal irradiation and 133 W/m² for direct normal irradiation with 5-minute resolution data, the model is about three times as accurate as the satellite weather prediction. This approach offers significant advantages in spatial resolution, reliability, and cost-effectiveness over conventional irradiance data by satellites or sensors. Utilizing SMARTBLUE AG'S dense network of thousands of monitored PV plants, the proposed methodology will enable the accurate prediction of irradiance in Germany, significantly enhancing asset management capabilities for PV plants.
In this work, the factors leading to string outages are examined, and an enhanced method for detecting faults at the sub-string level is presented. Utilizing GPT4-o to analyze O&M reports of 5089 photovoltaic plants, we classified outages according to the affected components and the underlying origin, identifying the most frequent string fault causes. An approach employing CUSUM Charts is introduced to identify substring outages within PV systems effectively. The methodology utilizes fundamental field data that is commonly available in practice. A filtering approach, combined with the use of CUSUM control charts, minimizes false positives, ensuring that only consistent underperformance is flagged as an out-age. The methodology returns far fewer false positives and more stable error intervals for substring outages than a former monitoring approach. Overall, the study demonstrates a significant improvement in detecting substring outages. The advanced methodology enables more effective O&M for PV plants, where substring outages are reliably identified after a short detection time.
