A researcher at Jönköping University in Sweden has developed an artificial intelligence-based monitoring system for photovoltaic panels using infrared thermography. This new methodology, which combines advanced machine learning techniques, stands out ability to resist disturbances such as rotation, noise, scaling and haze. The results? An accuracy of 98% in the training phase and 96.8% in testing, as explained by Dr. Waqas Ahmed, author of the study published on Energy Reports:
Current image processing-based approaches for health and fault diagnosis often have limitations related to specific datasets. These methods suffer from sensitivity to factors such as rotation, scaling, noise, blur and haze. Furthermore, deep learning algorithms, while powerful, require robust datasets and accurate hyperparameter optimization, as well as having high computational complexity and memory requirements.
How the photovoltaic panel monitoring system works
The innovative methodology involves the use of a infrared camera to capture thermographic images of the panels. These images, in the pre-processing process, are enhanced using a dehazing algorithm and contrast optimization. Subsequently, each image is divided into sub-images of 5×5 pixels.
Local feature extraction occurs through Gaussian and non-linear methods, eliminating redundant values and maintaining only 80% of the most significant information. Once the data is processed, a k-means clustering algorithm reduces the feature vector to 300 elements per image, optimizing memory use.
For model training they are used shallow classifierssuch as support vector machines (SVMs). The process employs 5-fold cross-validation to ensure effective training. The system then classifies the panels into three health categories: healthy, with hotspot or broken.
Comparison with other AI techniques
The system was tested on a 44.24 kW photovoltaic system consisting of 376 crystalline silicon modules of 240W each, located in Lahore, Pakistan. The images were captured under ambient conditions that included temperatures between 32°C and 40°C, wind speeds of 6.9 m/s and irradiance of 700 W/m². The infrared was split into datasets, using 80% of the images for training and 20% for validation.
The methodology has achieved impressive results, with a average accuracy of 96.8%accuracy values of 92%-100%, and F1-score of 0.958, 1.0 and 0.947 for the fault, healthy and hotspot classes respectively. Compared to other AI approaches, it showed similar or superior performance. For example, the RB-SIFT method achieved the best score with 98.66%, while other techniques such as SURF and pre-trained neural networks achieved 97%.
The Dr. Ahmed concludes that this method represents an important step forward in improving the reliability of photovoltaic panels, overcoming the memory and precision limitations of existing solutions.
