Dataset Source

The dataset was publicly released by the Sensing and Control Laboratory at Hangzhou Dianzi University.

References

Wu, S., et al., A feature space class balancing strategy-based fault classification method in solar photovoltaic modules. Engineering Applications of Artificial Intelligence, 2024. 136: p. 108991.https://doi.org/10.1016/j.engappai.2024.108991.

Data Collection

"The urban building rooftop photovoltaic dataset" is a deep learning dataset designed for studying photovoltaic systems installed on rooftops of urban buildings. We employed a Matrice 350 RTK UAV equipped with a Zenmuse H20 infrared imaging camera to capture a total of 1,724 thermographic images on the roofs of three building complexes. The samples encompassed six different classes. The UAV was positioned at a height of 8.5-10 meters above the PV panels at a vertical angle, resulting in clear images with a resolution of 128x200 pixels. The images were captured between 2-5 pm on a sunny day to ensure consistent lighting conditions.

Capture Conditions

• Equipment: Matrice 350 RTK UAV with Zenmuse H20 infrared imaging camera
• Number of Images: 1,724
• Location: Roofs of three building complexes
• Sample Classes: Six different classes
• Positioning: Height of 8.5-10 meters above the PV panels
• Time of Capture: Between 2-5 pm on a sunny day
• Image Resolution: 128x200 pixels

Definitions of different faults in the urban building rooftop photovoltaic dataset, along with possible manifestations and potential causes.

Number of Images for Each Class in the Urban Building Rooftop Photovoltaic Dataset

Here is the distribution of the number of images for each class in the dataset:

• Soiling: 835 images
• No-Anomaly: 319 images
• Dirty Spots: 236 images
• Shadowing: 201 images
• Soiled Shadows: 103 images
• Cell: 30 images