Abstract
Photovoltaic systems are a core component of eco-friendly energy technologies and are now widely utilized across the world for power generation. However, solar modules that are continuously exposed to the external environment experience gradual performance degradation, which results in significant power loss and operational problems. Existing aging diagnostic methods such as current-voltage curve analysis and electroluminescence/photoluminescence testing have limitations in terms of real-time monitoring, quantitative evaluation, and applicability to large-scale power plants. To address these challenges, this study proposes a novel degradation detection method that utilizes the first-order derivative of the voltage-power curve of solar modules to extract key features. This method can estimate the number of degraded solar modules within a string and the degree of degradation, enabling early detection of subtle changes in electrical characteristics. In this study, we developed an AI model based on long short-term memory to classify normal and abnormal states and predict aging status, thereby supporting monitoring and early diagnosis. The model architecture was designed to reflect the characteristics of solar power systems, adopting a relatively shallow network due to the time-series data not being excessively long and the feature changes being clear. This design effectively mitigates the issues of overfitting and gradient vanishing, thereby positively contributing to the stability of model training. The training and validation results of the proposed long short-term memory model were verified through MATLAB simulations, confirming its effectiveness in learning and convergence.