Development of a modified adaptive protection scheme using machine learning technique for fault classification in renewable energy penetrated transmission line
The conventional utility grid-protection scheme is predesigned at the network's early planning stage with consideration to the high short circuit fault current magnitude contribution level from the Synchronous Generators (SG) to prevent the mal-operation of the relaying scheme. In the modern po...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/92996/1/FK%202021%2012%20IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The conventional utility grid-protection scheme is predesigned at the network's early planning stage with consideration to the high short circuit fault current magnitude contribution level from the Synchronous Generators (SG) to prevent the mal-operation of the relaying scheme. In the modern power system grid, the integration of the Renewable Energy Resources (RER) from Windfarms (WF) or Photovoltaic (PV) generation sources focused on addressing the climate change environmental issues and solving the impending future energy sustainability challenges. In compliance with the new grid code requirement of permanently integrating RER with the conventional SG sources during grid short circuit faults, also known as the low voltage fault ride-through (LVFRT). Such RERs integration phenomenon compromised the existing protection relaying scheme operation settings due to the power grid system topology changes. The added infeed current penetration from integrated RERs impacted adversely on the existing protective relay operation setting compromise. The relay operation setting compromise is due to the wrong estimated impedance seen by the relay leading to overreach or underreach mal-operation. The current Adaptive Protection Scheme (APS) motivation focused on the accurate relay operation setting changes based on the prevailing grid system configuration variations. Hence, eliminate the utility grid relay operation setting compromise. The lack of healthy lines detailed protection useful information knowledge has limited the existing APS performance, as only faulty lines' measured parameters (voltage, current, and phase angle) are mostly used in the relaying protection scheme design. The high-cost of implementations, cyber-attack, and latency concerns from the adopted communication channels for the standard APS relay characteristic setting and selection is another drawback identified. This study proposed a modified standalone Machine Learning-based Adaptive Protection Scheme (ML-APS) relay' fault classifier model using novel useful hidden Knowledge Discovery from historical fault events Dataset (KDD) from healthy and faulty lines. The healthy lines extracted fault signals' functional signature are added to the earlier deployed faulty-line decomposed dataset, operation parameters, and changing network topology information from the SCADA logged reports without communication channel use. The hybrid Wavelet Multiresolution Analysis and Machine learning algorithm (WMRA-ML) is used to extracts the useful hidden knowledge from decomposed one-cycle fault transient signals (voltage & current) from four Matlab/Simulink CIGRE models. Consideration was given to different RER penetration levels based on the changing network topologies subjected to twelve different short circuit fault scenarios.The selected 29 unique feature attributes across 15,120 historical faults dataset deployed as the input-output training dataset for the ML-APS relay classifier model development in Waikato Environment of Knowledge Analysis Software (WEKA). The obtained result from the twelve deployed ML algorithms for the standalone intelligent ML-APS relay classifier modification without communication medium adoption for transmitting and receiving the updated relay operation settings during network configuration changes. The Random Tree standalone ML-AP relay model presented the best performing models from the ML-APS relay model with the best average performance for the correctly classified fault types of 97.61 % at 5 % significance level above other ML algorithms. The recorded kappa statistic value of 0.9802, and the Receiver Operating Curve (ROC) area of 98.73 %. The Random Tree relay algorithm model presented an improved average trip decision time of 18 ms compared with the standard minimum value of 20 ms recorded for the conventional relay due to eliminated communication channels. The ML-AP relay model addressed the cyber-attack and latency compromises in the earlier APS relay for the modern power system network. The obtained result demonstrated useful hidden knowledge in the healthy line sections that have contributed valuable information for improved ML-APS relay model for the faults detection, discrimination, and decision trip improvement during the grid short circuit faults. |
|---|