Fault diagnosis of uncertain photovoltaic systems using deep recurrent neural networks based Lissajous curves

Abstract

Data-driven approaches have gained significant interest in the fault detection and diagnosis (FDD) field, often utilizing numerous sensors for accurate and reliable monitoring. However, extensive sensor deployment can lead to increased costs, maintenance complexity, potential data redundancies, and uncertainties. This study proposes an innovative methodology to enhance model representation and improve decision-making processes by strategically reducing the number of sensors required, thereby addressing sensor-related challenges while maintaining effective fault diagnosis capabilities. The paper investigates the most prevalent experimental faults that can occur in grid-connected photovoltaic (GCPV) systems, such as sensor faults, PV panel faults, inverter faults, and grid connection faults, to ensure a thorough analysis of the system. Firstly, the number of required sensors is reduced. Then, Lissajous curves are applied to extract additional informative features, which are subsequently fed into deep learning classifiers; such as Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM); for fault diagnosis. Additionally, an extended approach based on interval-valued data representation is introduced to handle uncertainties, including measurement errors, noise, and variable variability. The methodology is experimentally validated using GCPV systems, comprehensively analyzing potential faults and their mitigation.

The results, demonstrated using noisy testing data, highlight the robustness and effectiveness of the proposed approach, achieving average accuracies of 94.36% and 99.50%. This confirms the approach’s capability to manage FDD challenges in PV systems, even under conditions that mimic real-world noise and uncertainties.