فهرست نشریات علمی دسترسی آزاد ایران

Energy is at the core of global challenges, influencing economic, environmental, and societal development. In this context, photovoltaic (PV) energy production has experienced unprecedented growth, driven by research advancements and technological progress. However, PV panels are susceptible to faults such as partial shading and short circuits, which reduce energy yield and compromise system reliability, underscoring the need for efficient fault diagnosis methods. This situation raises concerns regarding the effectiveness and dependability of PV system maintenance strategies. The objective of this study is to provide a comprehensive and quantitative benchmark of classical machine learning models for PV fault diagnosis, assessing both prediction accuracy and inference speed to facilitate real-time applications. Consequently, comparing machine learning models for fault diagnosis is essential. This study evaluates models based on two primary criteria: prediction accuracy and computational speed. The selected faults—partial shading and short circuits—were analyzed using dimensionality reduction techniques to balance class distributions and retain essential variables, thereby mitigating biased results. We assessed several machine learning models, including SVM (Support Vector Machine), KNN (K-Nearest Neighbors), MLP (Multi-Layer Perceptron), LR (Logistic Regression), DT (Decision Tree), and RF (Random Forest). Evaluation metrics included F1-score, recall, precision, accuracy, and confusion matrix analysis. The novelty of this study lies in the joint evaluation of accuracy and inference speed using a standardized synthetic dataset, providing a benchmark for real-time fault detection. Results indicate that the MLP model achieved the highest accuracy (100%) with an exceptional prediction time (~0.3118 milliseconds), whereas SVM demonstrated a favourable balance between speed and precision (99.85% accuracy, ~0.748 milliseconds). These findings highlight the superiority of MLP and RF models over other approaches and offer clear guidance for selecting models in real-time PV monitoring systems. As the literature reveals a notable gap in optimizing fault prediction time, this research contributes to the field by integrating both accuracy and computational efficiency, establishing a foundation for the development of real-time PV fault detection systems.