Fault Detection and Diagnosis of PV Systems Based on Corrected Performance Ratio Analysis and Adaptive Thresholds Using CNN-LSTM Hybrid Neural Network Prediction Model

Abstract

In the model-based approach, a trusted and reliable photovoltaic (PV) array model provides an essential baseline for monitoring the system’s health and diagnosing any underperformance. To this end, this article proposes a threephase methodology for developing a robust PV plant model and implementing a fault detection and diagnosis mechanism, suitable for large-scale grid-connected PV systems. The first stage involves constructing a single-layer hybrid CNNLSTM model aimed at predicting photovoltaic power generation. This model uses real-world measured data obtained from a large-scale, gridconnected photovoltaic system operating in Algeria. In this architecture, the convolution layer acts as a feature extractor, identifying local patterns in the measured data. Meanwhile, the LSTM component captures temporal dependencies. In the second phase, an automated inspection platform is developed, leveraging the evaluation of the Corrected Performance Ratio (CPR) for each inverter within the operational photovoltaic (PV) system... Next, a lower threshold and an upper threshold set by Hampel identifier are defined and processed, allowing the system to detect faults by generating a faulty signal. The diagnostic process is implemented by setting dynamic voltage and current thresholds derived from the prediction model. These thresholds are then adjusted adaptively to improve the sensitivity of fault detection for both voltage and current abnormalities. In the final phase, the proposed supervision system is implemented as a standalone graphical user interface (GUI) application based on the developed methodology. This validates its applicability for photovoltaic (PV) supervisory control and data acquisition (SCADA) systems. Evaluated case studies demonstrate the effectiveness of this approach in detecting typical PV systems faults, such as short circuits, string disconnections, and partial shading effects