<![endif]-->). Metode ANN digunakan untuk klasifikasi data DGA. Sedangkan metode PSO sebagai optimasi parameter ANN agar dapat bekerja lebih efektif. Tahapan penelitian meliputi preprocessing, penanganan data outlier, balancing data dengan SMOTE, optimasi parameter ANN dengan PSO dan evaluasi kinerja menggunakan nilai presisi dan akurasi model. Model PSO-ANN kemudian dibandingkan dengan metode ANN, Support Vector Machine (SVM), dan Random Forest Classifier (RFC). Hasil penelitian menunjukkan bahwa metode hybrid PSO-ANN dan RFC memberikan kinerja klasifikasi terbaik pada data DGA dengan akurasi dan presisi yang lebih unggul dibandingkan metode lainnya. Metode PSO-ANN dan RFC mencapai kinerja terbaik dengan rata-rata presisi diatas 0,90, mengungguli ANN 0,82 dan SVM 0,80. PSO-ANN unggul dalam klasifikasi gangguan termal (T1-T2) dan low energy discharge (D1), meskipun masih terdapat tantangan dalam mendeteksi kelas D2 dan T3 akibat keterbatasan data minoritas. Pendekatan metode yang digunakan efektif mengatasi permasalahan ketidakseimbangan data dan mampu meningkatkan ketepatan diagnosis jenis gangguan trafo. Implikasi penelitian ini memberikan solusi diagnostik yang lebih akurat untuk pemeliharaan preventif trafo, dengan rekomendasi peningkatan dataset dan eksplorasi teknik optimasi lebih lanjut untuk kelas minoritas. PT. PLN (Persero) manages highly complex and diverse transmission assets. One of the most critical components of transmission assets is the power transformer. Transformers play a key role in the distribution of electrical energy to consumers, and their reliability depends heavily on the condition of insulating oil. Gas concentration in transformer insulation oil is used as a method for detecting the first internal transformer failures known as dissolving gas analysis (DGA). However, the conventional interpretation of DGA data is often inaccurate due to the complexity of the gas distribution and data imbalance. Interpretation techniques such as the Doernenburg Ratio, the Duval Triangle or the IEC 60599 standards often produce inconsistent and overlapping results between different types of defects. In addition, manual interpretation is often subjective and requires experts with in-depth knowledge of the gas characteristics in DGA test results. Therefore, this research doing comparative study to classifies the types of transformer faults with DGA data using a hybrid method combining Particle Swarm Optimization (PSO) and Artificial Neural Network (ANN). The ANN method is used for DGA data classification. Meanwhile, the PSO method serves as an optimization of ANN parameters to allow it to work more effectively. The research stages include preprocessing, handling outlier data, balancing data with SMOTE, optimization of ANN parameters with PSO and performance evaluation in terms of model precision and accuracy. The PSO-ANN model is then compared to the ANN, the support vector machine (SVM) and the random forest classifier (RFC) method. The results show that hybrid method PSO-ANN and RFC provide the best classification performance on DGA data with superior accuracy and precision compared to other methods. The PSO-ANN and RFC methods achieved the best performance with an average precision above 0.90, outperforming ANN (0.82) and SVM (0.80). PSO-ANN excelled in classifying thermal faults (T1-T2) and low energy discharge (D1), although challenges remain in detecting classes D2 and T3 due to the limitations of minority data. The methods employed effectively address data imbalance issues and enhance the accuracy of transformer fault type diagnosis. The implications of this study provide more accurate diagnostic solutions for preventive transformer maintenance, with recommendations for improving the dataset and exploring further optimization techniques for minority classes."> <![endif]-->). Metode ANN digunakan untuk klasifikasi data DGA. Sedangkan metode PSO sebagai optimasi parameter ANN agar dapat bekerja lebih efektif. Tahapan penelitian meliputi preprocessing, penanganan data outlier, balancing data dengan SMOTE, optimasi parameter ANN dengan PSO dan evaluasi kinerja menggunakan nilai presisi dan akurasi model. Model PSO-ANN kemudian dibandingkan dengan metode ANN, Support Vector Machine (SVM), dan Random Forest Classifier (RFC). Hasil penelitian menunjukkan bahwa metode hybrid PSO-ANN dan RFC memberikan kinerja klasifikasi terbaik pada data DGA dengan akurasi dan presisi yang lebih unggul dibandingkan metode lainnya. Metode PSO-ANN dan RFC mencapai kinerja terbaik dengan rata-rata presisi diatas 0,90, mengungguli ANN 0,82 dan SVM 0,80. PSO-ANN unggul dalam klasifikasi gangguan termal (T1-T2) dan low energy discharge (D1), meskipun masih terdapat tantangan dalam mendeteksi kelas D2 dan T3 akibat keterbatasan data minoritas. Pendekatan metode yang digunakan efektif mengatasi permasalahan ketidakseimbangan data dan mampu meningkatkan ketepatan diagnosis jenis gangguan trafo. Implikasi penelitian ini memberikan solusi diagnostik yang lebih akurat untuk pemeliharaan preventif trafo, dengan rekomendasi peningkatan dataset dan eksplorasi teknik optimasi lebih lanjut untuk kelas minoritas. PT. PLN (Persero) manages highly complex and diverse transmission assets. One of the most critical components of transmission assets is the power transformer. Transformers play a key role in the distribution of electrical energy to consumers, and their reliability depends heavily on the condition of insulating oil. Gas concentration in transformer insulation oil is used as a method for detecting the first internal transformer failures known as dissolving gas analysis (DGA). However, the conventional interpretation of DGA data is often inaccurate due to the complexity of the gas distribution and data imbalance. Interpretation techniques such as the Doernenburg Ratio, the Duval Triangle or the IEC 60599 standards often produce inconsistent and overlapping results between different types of defects. In addition, manual interpretation is often subjective and requires experts with in-depth knowledge of the gas characteristics in DGA test results. Therefore, this research doing comparative study to classifies the types of transformer faults with DGA data using a hybrid method combining Particle Swarm Optimization (PSO) and Artificial Neural Network (ANN). The ANN method is used for DGA data classification. Meanwhile, the PSO method serves as an optimization of ANN parameters to allow it to work more effectively. The research stages include preprocessing, handling outlier data, balancing data with SMOTE, optimization of ANN parameters with PSO and performance evaluation in terms of model precision and accuracy. The PSO-ANN model is then compared to the ANN, the support vector machine (SVM) and the random forest classifier (RFC) method. The results show that hybrid method PSO-ANN and RFC provide the best classification performance on DGA data with superior accuracy and precision compared to other methods. The PSO-ANN and RFC methods achieved the best performance with an average precision above 0.90, outperforming ANN (0.82) and SVM (0.80). PSO-ANN excelled in classifying thermal faults (T1-T2) and low energy discharge (D1), although challenges remain in detecting classes D2 and T3 due to the limitations of minority data. The methods employed effectively address data imbalance issues and enhance the accuracy of transformer fault type diagnosis. The implications of this study provide more accurate diagnostic solutions for preventive transformer maintenance, with recommendations for improving the dataset and exploring further optimization techniques for minority classes.">
KLASIFIKASI KERUSAKAN TRAFO MENGGUNAKAN METODE PARTICLE SWARM OPTIMIZATION (PSO) DAN ARTIFICIAL NEURAL NETWORK (ANN)
Yusuf Irkham, Teguh Bharata Adji, S.T., M.T., M.Eng., Ph.D. ; Ir. Azkario Rizky Pratama, S.T., M.Eng., Ph.D.
2025 | Tesis | S2 Teknologi Informasi
PT. PLN (Persero) mengelola aset transmisi yang sangat kompleks dan
beragam. Salah satu peralatan yang sangat penting bagi aset transmisi adalah
trafo tenaga. Trafo memiliki peran yang vital dalam penyaluran energi listrik
kepada konsumen dan keandalannya sangat bergantung pada kondisi minyak isolasi.
Konsentrasi gas dalam minyak isolasi trafo digunakan sebagai metode untuk
mendeteksi secara dini gangguan internal trafo yang disebut Dissolved Gas
Analysis (DGA). Namun, interpretasi data DGA secara konvensional kurang
akurat karena kompleksitas distribusi gas dan ketidakseimbangan data. Teknik
interpretasi seperti Doernenburg Ratio, Duval Triangle atau standar IEC 60599
menunjukkan hasil yang tidak konsisten dan tumpang tindih antarjenis gangguan. Selain
itu, interpretasi manual seringkali subyektif dan memerlukan ahli dengan
pemahaman mendalam tentang karakteristik gas-gas dalam hasil uji DGA.
Oleh karena itu, penelitian ini melakukan studi komparasi klasifikasi jenis
kerusakan trafo dengan data DGA menggunakan metode hybrid antara lain Particle
Swarm Optimization (PSO) dan Artificial Neural Network (ANN<!--[if supportFields]>
XE "ANN" <![endif]--><!--[if supportFields]><![endif]-->). Metode ANN digunakan untuk klasifikasi data DGA. Sedangkan metode PSO
sebagai optimasi parameter ANN agar dapat bekerja lebih efektif. Tahapan
penelitian meliputi preprocessing, penanganan data outlier, balancing
data dengan SMOTE, optimasi parameter ANN dengan PSO dan evaluasi
kinerja menggunakan nilai presisi dan akurasi model. Model PSO-ANN kemudian
dibandingkan dengan metode ANN, Support Vector Machine (SVM), dan Random
Forest Classifier (RFC).
Hasil penelitian menunjukkan bahwa metode hybrid PSO-ANN dan RFC
memberikan kinerja klasifikasi terbaik pada data DGA dengan akurasi dan presisi
yang lebih unggul dibandingkan metode lainnya. Metode PSO-ANN dan RFC mencapai
kinerja terbaik dengan rata-rata presisi diatas 0,90, mengungguli ANN 0,82 dan
SVM 0,80. PSO-ANN unggul dalam klasifikasi gangguan termal (T1-T2) dan low
energy discharge (D1), meskipun masih terdapat tantangan dalam mendeteksi kelas
D2 dan T3 akibat keterbatasan data minoritas. Pendekatan metode yang digunakan
efektif mengatasi permasalahan ketidakseimbangan data dan mampu meningkatkan
ketepatan diagnosis jenis gangguan trafo. Implikasi penelitian ini memberikan
solusi diagnostik yang lebih akurat untuk pemeliharaan preventif trafo, dengan
rekomendasi peningkatan dataset dan eksplorasi teknik optimasi lebih lanjut
untuk kelas minoritas.
PT. PLN (Persero) manages highly complex and diverse transmission
assets. One of the most critical components of transmission assets is the power
transformer. Transformers play a key role in the distribution of electrical
energy to consumers, and their reliability depends heavily on the condition of
insulating oil. Gas concentration in transformer insulation oil is used as a
method for detecting the first internal transformer failures known as
dissolving gas analysis (DGA). However, the conventional interpretation of DGA
data is often inaccurate due to the complexity of the gas distribution and data
imbalance. Interpretation techniques such as the Doernenburg Ratio, the Duval
Triangle or the IEC 60599 standards often produce inconsistent and overlapping
results between different types of defects. In addition, manual interpretation
is often subjective and requires experts with in-depth knowledge of the gas
characteristics in DGA test results.
Therefore, this research doing comparative study to classifies the types
of transformer faults with DGA data using a hybrid method combining Particle
Swarm Optimization (PSO) and Artificial Neural Network (ANN). The ANN method is
used for DGA data classification. Meanwhile, the PSO method serves as an
optimization of ANN parameters to allow it to work more effectively. The research
stages include preprocessing, handling outlier data, balancing data with SMOTE,
optimization of ANN parameters with PSO and performance evaluation in terms of
model precision and accuracy. The PSO-ANN model is then compared to the ANN,
the support vector machine (SVM) and the random forest classifier (RFC) method.
The results show that hybrid method PSO-ANN and RFC provide the best
classification performance on DGA data with superior accuracy and precision
compared to other methods. The PSO-ANN and RFC methods achieved the best
performance with an average precision above 0.90, outperforming ANN (0.82) and
SVM (0.80). PSO-ANN excelled in classifying thermal faults (T1-T2) and low
energy discharge (D1), although challenges remain in detecting classes D2 and
T3 due to the limitations of minority data. The methods employed effectively
address data imbalance issues and enhance the accuracy of transformer fault
type diagnosis. The implications of this study provide more accurate diagnostic
solutions for preventive transformer maintenance, with recommendations for
improving the dataset and exploring further optimization techniques for
minority classes.
Kata Kunci : klasifikasi, dissolved gas analysis, minyak trafo, particle swarm optimization, artificial neural network, support vector machine, random forest classifier