Loading...
Faults detection and classifications for grid connected PV system using deep learning methods
Author(s)
Makosso, Thomas Lionel
Date Issued
2026
Type
doctoral thesis
Publisher
Cape Peninsula University of Technology
Abstract
This study provides a thorough approach for fault detection and classification in a 200 kW gridconnected photovoltaic (PV) system using Long Short-Term Memory (LSTM) neural networks. As PV systems and other renewable energy sources are being included into modern power networks, stability and reliability must be maintained. One of the biggest problems is recognizing and classifying short-circuit failures, which can result in serious damage and downtime if ignored. Conventional fault detection techniques frequently fall short in addressing their limitations in practical settings due to their inability to adjust to a variety of fault types and system dynamics. Traditional methods struggle with detection accuracy because system dynamics and fault types vary greatly. LSTM networks, which are great at managing timeseries data and detecting temporal relationships in electrical signals, are used in the proposed method. By analysing voltage and current waveforms, the LSTM model effectively detects and classifies a wide variety of short-circuit fault types, including single-line-to-ground, line-to-line, and three-phase faults. The grid-connected PV system is simulated using Python, and synthetic fault data is generated for the training and validation of the LSTM network. By creating a solid dataset with a variety of fault scenarios, this study also fills an empirical gap and allows for a more thorough assessment of the model's performance. The model is trained on a large dataset that includes both usual operating conditions and different fault scenarios. Data augmentation approaches, like the Synthetic Minority Over-Sampling Technique (SMOTE), are also utilized to address the imbalance in the fault data so that the model can accurately categorize less prevalent problem types. The effectiveness of the LSTM-based fault detection and classification system is evaluated using confusion matrix measurements and receiver over the curve (ROC). The simulation's results demonstrate that the LSTM model outperforms conventional techniques in terms of accuracy, defect detection rates, and response times. In contrast to MATLAB, which offered configured blocks that were typically set to predict the best performance and did not permit investigation because poor results scenarios could not be evaluated and upgraded as the study that was conducted, Python allowed for a great understanding of the development of a deep learning model.
Additional information
Thesis (DEng (Electrical Engineering))--Cape Peninsula University of Technology, 2026
File(s)![Thumbnail Image]()
Loading...
Name
Makosso, TL_218283695.pdf
Size
3.72 MB
Format
Adobe PDF
Checksum
(MD5):e988bc48a18418874897beeafa99752d