Time-Frequency Image Representation Aided Deep Feature Extraction-Based Grid Connected Solar PV Fault Classification Framework

Accurate detection of faults in grid connected solar PV systems is important to ensure the reliability of power systems with distributed generation. Considering the aforesaid fact, here, a smoothed pseudo-Wigner-Ville distribution (SPWVD) and stacked sparse autoencoder (SSA) based automated feature extraction technique is proposed for accurate detection of faults in grid connected solar PV systems. To this end, three phase current data of normal as well as different fault scenarios obtained from point of common coupling (PCC) were converted into direct (d) and quadrature (q) axis using extended Park’s vector approach. Then, the obtained d-axis (I_d) and q-axis (I_q) currents were converted to 2D time-frequency images using SPWVD. The converted time-frequency spectrum of the normal as well as faulty current data were used as inputs to the proposed SSA model for deep feature extraction. After extraction of deep features using SSA, analysis of variance (ANOVA) test and false discovery rate (FDR) correction was employed to select the most discriminative features. The feature selection was followed by classification using machine learning classifiers. It has been observed that the proposed technique achieved mean fault recognition accuracy of 98.79 and 97.56% for d-axis and q-axis currents respectively, respectively. The present approach can be used for accurate diagnosis of faults in grid connected solar PV systems.