Robust diagnosis of inverter faults and grid anomalies in photovoltaic systems using ensemble-based classification

Abstract

The increasing reliance on grid-connected photovoltaic (PV) systems for sustainable energy underscores the need for effective fault diagnosis to ensure system reliability and performance. This paper proposes a novel methodology that transforms multi-channel one-dimensional time series signals from PV systems into two-dimensional images, enabling feature extraction using a pre-trained MobileNetV1 network. We evaluate various classifiers, including gradient boosting, K-nearest neighbors, support vector machine, and a proposed voting classifier (VCL) that combines these models to improve accuracy and robustness. The dataset comprises 1,800 images across three classes: healthy systems, inverter faults, and grid anomalies, including tests on noisy and unbalanced data distributions to assess practical applicability. Experimental results show that the VCL outperforms individual classifiers and existing state-of-the-art methods, achieving high performance across all metrics with an overall accuracy of 99.81%. These findings demonstrate the effectiveness of the proposed approach for reliable and robust PV system fault diagnosis.