Heuristically Optimized Features Based Machine Learning Technique for Identification and Classification of Faults in PV Array

The faults in photovoltaic (PV) array lead to increased system losses and even fire hazards. The most frequent faults in PV strings are line-to-line (LL) and line-to-ground (LG) faults. Many efforts have been made to develop machine learning-based methods that are capable of detecting faults. However, these methods do not consider low mismatch faults, high impedance faults, active MPPT control, the effect of blocking diodes, step changes in irradiation levels and partial shading conditions in a single window. In this article, a novel and efficient modified binary genetic algorithm (MBGA) based on the weighted K-nearest neighbor method, which incorporates all the abovementioned constraints, has been proposed to identify and classify faults. In addition, it also gives information about the severity of faults. Unlike other machine learning (ML)-based methods, the developed technique considers features based on both frequency and time domain and employs MBGA to extract the optimal set of features, which further improves the accuracy of the algorithm and reduces the size of the dataset. The proposed method efficiently distinguishes faults from sudden shading conditions as both have similar characteristics and prevent false detection. Moreover, it has been verified that the developed method detects faults with an accuracy of 97.3% and classifies LL and LG faults with a precision of 99.25%.

I. Introduction

With the fast-growing population and enhancement in technology, energy demand has escalated in recent years as conventional energy sources are inadequate to accomplish the enormous needs for energy. This dearth of energy shifts our attention toward renewable energy sources owing to their accessibility and climate-friendly nature. Solar energy is the most promising and cost-effective solution for power generation among all nonconventional sources. A recent report by National Renewable Energy Laboratory (NREL) [1] depicts that photovoltaic (PV)-based generation would suffice 40 to 45% of the nation's electricity demand by 2050. However, over the last few years, it was found that losses owing to incipient PV faults range from 4% to 18% of the rated capacity, impacting the panel lifespan and efficiency, sometimes leading to fire hazards. This reflects the inability of the traditional protection scheme to identify faults in the PV array. The conventional methods are incompetent in detecting low mismatch faults, faults with high impedance, low irradiance faults, and faults under shading conditions. The detection of faults becomes more difficult in the presence of the active maximum power point tracking (MPPT) control and blocking diode because of the similar transient in terminal parameters due to fault events and variations in environmental conditions. Hence, it becomes very challenging to identify associated faults and discriminate them from temporary faults by using traditional protection schemes. This motivates researchers to propose a new, reliable, and efficient PV array fault detection technique.