Anomaly Detection in Photovoltaic Systems Using Improved Independent Component Analysis

Reliable operation of photovoltaic (PV) systems requires effective monitoring strategies to detect sensor faults that may degrade performance or compromise safety. This paper proposes an enhanced data-driven fault detection framework that combines Improved Independent Component Analysis (ICA) with the Kantorovich Distance (KD) and Kernel Density Estimation (KDE) for robust anomaly detection. ICA is employed to model multivariate dependencies among PV sensor measurements, such as irradiance, temperature, and power output, enabling the extraction of residuals that isolate abnormal variations. The KD metric is then used to quantify distributional shifts in the residuals between healthy and testing conditions, capturing both mean and covariance changes. To enable adaptive and non-parametric thresholding, KDE is applied to the KD values computed under normal conditions, facilitating statistically grounded fault detection. The proposed ICA–KD–KDE framework is validated on PV datasets with injected sensor faults, including bias, drift, and intermittent errors. Experimental results demonstrate superior sensitivity and low false alarm rates compared to conventional ICA-based methods, particularly in detecting subtle anomalies in environmental sensors. This approach provides a flexible and interpretable monitoring solution for real-time PV system diagnostics.