Resumen de Employing artificial intelligence techniques for the estimation of energy production in photovoltaic solar cells based on electroluminescence images
Héctor Felipe Mateo Romero
The investment in renewable energies has increased significantly in recent years, with photovoltaic solar energy emerging as one of the most prominent sources. The shift towards sustainable energy solutions is driven by the urgent need to mitigate climate change, reduce greenhouse gas emissions, and achieve energy independence. Photovoltaic (PV) technology, which converts sunlight directly into electricity using semiconductor materials, has seen substantial advancements in efficiency and cost reduction, making it a viable and attractive option for large-scale energy production.
Traditionally, the maintenance of photovoltaic plants has relied on manual labor to inspect the conditions of numerous solar modules. This involves technicians physically examining each module, looking for defects such as cracks, hot spots, and dirt accumulation that can impair performance. While effective for small installations, this manual approach becomes impractical for large-scale installations comprising thousands or even millions of modules. The sheer scale of these operations requires a more efficient and comprehensive method to ensure optimal performance and longevity of the PV systems.
The integration of artificial intelligence in this field has offered a pathway to optimize production and maintenance. AI technologies, including machine learning and computer vision, can automate the inspection process, providing rapid and accurate detection of defects. Presently, detecting problems on the surface of photovoltaic cells involves identifying patterns of defects using various technologies such as infrared thermography, electroluminescence imaging, and aerial drone inspections. However, this method has limitations because it does not directly correlate these issues with energy production, leading to a potential gap in understanding the actual impact of detected defects on the overall performance of the PV modules.
This thesis proposes a novel approach to analyzing the state of photovoltaic cells, serving as the initial step toward developing a system applicable to entire modules. The analysis involves developing an AI model capable of estimating the energy production of a photovoltaic cell using its electroluminescence image. Electroluminescence imaging is a powerful diagnostic tool that can reveal otherwise invisible defects by capturing the emitted light when a current is applied to the cell. By correlating these images with the cells performance data, specifically the current-voltage (I-V) curves, the AI model can predict the energy output and identify cells that may underperform due to defects.
The research has led to the creation of several datasets featuring various types of photovoltaic cells, encompassing different technologies and defect types. These datasets are crucial for training and validating the AI models, ensuring they can generalize across different conditions and cell types. Different proposals have been explored to address the problem, analyzing the advantages and disadvantages of each. For instance, various machine learning algorithms, including Convolutional Neural Networks or Random Forest. These algorithms have been tested to determine the most effective approach for image-based energy prediction.
Additionally, secondary achievements of the thesis include analyzing data imbalance using synthetic datasets and investigating the issue of series resistance observed in various cells. Data imbalance can bias the AI model, leading to poor performance on rare defects. By generating synthetic datasets, the research mitigates this issue, ensuring a balanced representation of all defect types. Series resistance, which affects the flow of current through the cell, is another critical parameter influencing performance. The investigation into series resistance provides insights into its impact on energy production and how it can detected with the EL images.
This thesis contributes by introducing an innovative AI-based approach for the precise estimation of energy production from electroluminescence images. This work not only enhances the efficiency of maintenance and monitoring of PV cells but also sets the stage for implementing it at module level or even in large-scale PV installations.
