Evaluasi Efektivitas Neural Network Classifier Dan Feedforward Neural Network Untuk Deteksi Arcing Seri Pada Sistem DC

Masadah, Richinta Tanisa (2026) Evaluasi Efektivitas Neural Network Classifier Dan Feedforward Neural Network Untuk Deteksi Arcing Seri Pada Sistem DC. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5022221205_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Peningkatan pemanfaatan sistem kelistrikan arus DC, khususnya pada sistem fotovoltaik (PV), menuntut adanya sistem proteksi yang mampu mendeteksi gangguan arcing seri secara cepat dan akurat. Gangguan arcing seri berbahaya karena dapat menghasilkan panas tinggi dan memicu kebakaran, namun sulit dideteksi oleh perangkat proteksi konvensional karena arus gangguannya relatif kecil, tidak stabil, dan tidak selalu menyebabkan kenaikan arus yang signifikan. Penelitian ini bertujuan mengevaluasi efektivitas Neural Network Classifier dan Feedforward Neural Network (FNN) dalam mengklasifikasikan empat kondisi operasi arus DC, yaitu Normal, Transition Start, During Arc, dan Transition End. Data arus diperoleh dari pengujian laboratorium menggunakan Power Supply Unit (PSU) sebagai sumber DC, kemudian diproses melalui windowing, flagging, ekstraksi fitur menggunakan Fast Fourier Transform (FFT), seleksi fitur berbasis Cross Correlation, normalisasi, serta pemodelan supervised learning. Dataset pemodelan terdiri atas 600 data yang dibagi menjadi 70% data training, 15% data validation, dan 15% data testing. Hasil pengujian menunjukkan bahwa Neural Network Classifier memperoleh akurasi training sebesar 99,29%, validation sebesar 98,89%, testing sebesar 98,89%, dan akurasi keseluruhan sebesar 99,17%. Sementara itu, FNN memperoleh akurasi training sebesar 97,86%, validation sebesar 98,89%, testing sebesar 95,56%, dan akurasi keseluruhan sebesar 97,67%. Pada data testing, Neural Network Classifier hanya menghasilkan satu kesalahan klasifikasi dengan gap training-testing sebesar 0,40%, sedangkan FNN menghasilkan empat kesalahan klasifikasi dengan gap training-testing sebesar 2,30%. Berdasarkan akurasi, jumlah kesalahan klasifikasi, dan kestabilan generalisasi, Neural Network Classifier dinilai lebih efektif dibandingkan FNN untuk deteksi arcing seri pada sistem DC berbasis fitur FFT.
==========================================================================================================================
The increasing use of direct current (DC) electrical systems, particularly photovoltaic (PV) systems, requires a protection system capable of detecting series arc faults rapidly and accurately. A series arc fault is hazardous because it can generate high temperature and potentially trigger fire; however, it is difficult to detect using conventional protection devices because the fault current is relatively small, unstable, and does not always cause a significant current increase. This study evaluates the effectiveness of a Neural Network Classifier and a Feedforward Neural Network (FNN) in classifying four DC current operating conditions, namely Normal, Transition Start, During Arc, and Transition End. The current data were obtained from laboratory testing using a Power Supply Unit (PSU) as the DC source and processed through windowing, flagging, feature extraction using Fast Fourier Transform (FFT), Cross Correlation-based feature selection, normalization, and supervised learning classification. The modeling dataset consisted of 600 balanced samples divided into 70% training data, 15% validation data, and 15% testing data. The results show that the Neural Network Classifier achieved training, validation, testing, and overall accuracies of 99.29%, 98.89%, 98.89%, and 99.17%, respectively. Meanwhile, the FNN achieved training, validation, testing, and overall accuracies of 97.86%, 98.89%, 95.56%, and 97.67%, respectively. In the testing dataset, the Neural Network Classifier produced only one misclassification with a training-testing gap of 0.40%, while the FNN produced four misclassifications with a training-testing gap of 2.30%. Based on accuracy, classification errors, and generalization stability, the Neural Network Classifier is more effective than the FNN for FFT-based series arc fault detection in DC systems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Arcing seri DC, Fast Fourier Transform, Cross Correlation, Neural Network Classifier, Feedforward Neural Network. DC series arcing, Fast Fourier Transform, Cross Correlation, Neural Network Classifier, Feedforward Neural Network.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK3401 Insulation and insulating materials T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5102.9 Signal processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User:	Richinta Tanisa Masadah
Date Deposited:	24 Jul 2026 01:08
Last Modified:	24 Jul 2026 01:08
URI:	http://repository.its.ac.id/id/eprint/137098

Actions (login required)

View Item