Logically Optimized and Probabilistic Integrated Photovoltaic Fault Finding Package based on Machine Learning

Peyman Ghaedi¹Aref Eskandari²Amir Nedaie¹Farzad Hatami¹Parviz Parvin¹Mohammadreza Aghaei^3*

• ¹Amirkabir University of Technology, Tehran, Iran
• ²Iran University of Science and Technology, Tehran, Iran
• ³NTNU, Trondheim, Norway

Artificial intelligence (AI) has been vastly utilized as an effective solution to detect photovoltaic (PV) faults and failures especially those which are most likely to remain undetected due to specific weaknesses in conventional protection devices. Nevertheless, multiple challenges can be witnessed in previous AI-based studies which can be listed as (i) neglecting critical fault detection conditions, (ii) using massive complex datasets with considerable dimensions, (iii) presenting models which are unable to detect and classify various faults in PV arrays both simultaneously and accurately. To address the mentioned challenges, this paper proposes a novel approach based on fuzzy logic (FL) system and particle swarm optimization (PSO) algorithm for detect and classification of several faults in PV arrays under various environmental and electrical conditions. To this end, initial dataset is constructed using the PV array current-voltage (I-V) characteristic curve. Using the initial attributes, numerous features based on Manhattan distance (MD), and Chebyshev distance (CD) are extracted. A larger group of machine learning (ML) classifiers is first considered. The FL system then nominates the most reliable classifiers based on accuracy, F1-score, and standard deviation, and subsequently, the PSO algorithm determines the optimal subset to maximize accuracy while minimizing standard deviation simultaneously. To produce a more accurate result, each classifier is weighted according to its accuracy. For this purpose, PSO is once again employed to determine the optimal weights of finally selected classifiers. Finally, various output prediction combining techniques are presented to provide the most accurate final prediction. An experiment is carried out on model verification using a dataset including PV array normal conditions as well as line-to-line (LL), open-circuit (OC), and degradation (DEG) faults in PV arrays under various environmental (irradiance and temperature) and electrical (mismatch and impedance) conditions. The results indicate that the proposed FL and PSO based model yields outstanding accuracy in detecting and classifying faults in PV arrays. The proposed method is then compared to recent studies which illustrates its outperformance over all other methods.