Recurrence Plot and Change Quantile-based Deep Supervised and Semi-supervised Protection for Transmission Lines Connected to Photovoltaic Plants

TL;DR

This paper introduces a novel deep learning-based protection scheme for transmission lines connected to photovoltaic plants, utilizing recurrence plots, change quantiles, and semi-supervised learning to improve fault detection and localization under various conditions.

Contribution

It proposes a new single-ended protection method combining recurrence matrices, InceptionTime, and semi-supervised learning to enhance fault detection in converter-based energy sources.

Findings

Effective fault detection across various system configurations.

Robust performance under noise and partial data labeling.

Accurate fault localization and phase identification.

Abstract

Conventional relays encounter difficulties in protecting transmission lines (TLs) connected to converter-based energy sources (CBESs) due to the influence of power electronics on fault characteristics. This article proposes a single-ended intelligent protection method for the TL segment between the grid and a Photovoltaic (PV) plant. The approach utilizes a Recurrence Matrix and an InceptionTime-based system to identify faults by using the mean change in quantiles of 3-phase currents. It determines the fault position and identifies the faulty phase. ReliefF feature selection is applied to extract the optimal quantile features. The scheme's performance is assessed under abnormal conditions, including faults and capacitor and load-switching events, simulated in Power Systems Computer Aided Design / Electromagnetic Transients Program (PSCAD/EMTDC) on the Western System Coordinating Council (WSCC) 9-bus system, with various fault and switching parameters. The scheme is also validated on the New England IEEE 39-bus system and in presence of partially rated converters. Additionally, the validation of the proposed strategy takes into account various conditions, including double-circuit line configuration, noise, series compensation, high-impedance faults, current transformer (CT) saturation, evolving and cross-country faults, remote and local faults, as well as variations in PV capacity, sampling frequency, and data window size. To address label scarcity and improve generalization, semi-supervised learning paradigms including label spreading, label propagation, and self-training are integrated with the InceptionTime framework, enabling near-supervised performance with limited annotated fault data. The results demonstrate that the approach is effective in handling different system configurations and conditions, ensuring the protection of TLs connected to large PV plants.