Sensor fault detection and diagnosis in photovoltaic systems using Hellinger Distance and Individual Conditional Expectation analysis

Accurate fault detection in photovoltaic (PV) systems is crucial for maintaining optimal performance and extending the system lifespan. This paper introduces a sensor fault detection and diagnosis approach that combines Partial Least Squares (PLS) regression with a Hellinger Distance (HD)-based monitoring chart, offering advantages in handling multicollinearity and reducing dimensionality in high-dimensional datasets. PLS generates residuals that capture deviations, which are then analyzed using an HD-based monitoring chart with thresholds determined by Kernel Density Estimation (KDE). This method enhances sensitivity to various sensor faults, including bias, drift (or aging), and intermittent faults. In this study, sensor faults were injected into both pyranometers and temperature sensors to validate the effectiveness of the proposed approach. In this context, weather data and solar irradiance serve as input variables in the PLS model, while PV power production is the output variable. Compared to conventional PLS monitoring charts, the PLS-HD approach demonstrates superior performance in fault detection. For fault diagnosis, Individual Conditional Expectation (ICE) plots are utilized to investigate the impact of faults by visualizing changes in the relationship between input features and model predictions under faulty conditions. This integrated approach provides a robust and interpretable framework for detecting and diagnosing sensor faults in PV systems.