PV string failures prediction through machine learning analytics

A challenging task for system owners and PV power plant operators is deploying the plant's reliability. This study discloses and reports the system malfunction to the system owners and plant operators, supporting effective planning of preventive maintenance and operating of the system, reducing downtime and improving economic aspects. Using different Machine Learning (ML) algorithms, namely support vector machine (SVM), K-nearest neighbors (KNN), Decision Tree (DT), and multi-layer perceptron (MLP), classifying failure and normal using one year of data from an operating 9.5 MWp PV system in central Thailand. Data was labelled anomalies by comparing normalising string currents in a junction box. Supervised ML algorithms used the labelled dataset for training and testing models as 80% and 20%, respectively. In addition, the dataset was split into three groups, 3, 6, and 12 months, analysing the effects of the number of datasets on the performance of ML. The results indicated that multi-layer perception (MLP) presented the highest modelling performance, R2, at approximately 90 per cent 30 minutes in advance using a full-year historical dataset. While the Decision tree (DT) showed the lowest performance, below 65 per cent, in all ranges of the dataset, which mainly influences the performance of ML