Identifikasi Parameter Dominan terhadap Kondisi Corrosive Sulfur pada Minyak Isolasi Transformator Berbasis Data Analisis Minyak

RAHMAT FITRIADI, Prof. Ir. Tumiran, M. Eng., Ph. D., IPU.; Syukron Abu Ishaq Alfarozi, S.T., Ph.D.

2026 | Tesis | S2 Teknik Elektro

Transformator tenaga merupakan aset kritis dalam sistem tenaga listrik yang keandalannya dipengaruhi oleh kondisi minyak isolasi transformator. Salah satu fenomena degradasi yang perlu diperhatikan adalah corrosive sulfur, yaitu kondisi ketika senyawa sulfur reaktif dalam minyak bereaksi dengan material tembaga dan berpotensi membentuk deposit tembaga sulfida (Cu2S). Deposit tersebut dapat menurunkan kekuatan dielektrik sistem isolasi minyak--kertas dan meningkatkan risiko kegagalan transformator. Penelitian ini bertujuan untuk mengidentifikasi parameter oil analysis yang memiliki kontribusi prediktif dominan dalam membedakan kondisi corrosive dan non-corrosive, serta menganalisis peran dan tingkat keterkendalian operasional parameter tersebut.

Dataset penelitian terdiri atas 137 data historis dari enam transformator tenaga selama periode 2012--2025 yang mencakup parameter fisika-kimia minyak, Dissolved Gas Analysis (DGA), dan Furan (2FAL). Tahapan penelitian meliputi seleksi data, pengkodean target biner, Exploratory Data Analysis (EDA), evaluasi hubungan antara parameter numerik dan target biner menggunakan korelasi point-biserial, serta pemodelan klasifikasi menggunakan Logistic Regression, Random Forest, dan XGBoost. Imputasi nilai hilang dan transformasi data dilakukan di dalam pipeline pemodelan untuk mencegah data leakage. Performa model dievaluasi menggunakan ROC-AUC pada data uji, F1-score dengan kelas corrosive sebagai kelas positif, serta rata-rata dan standar deviasi ROC-AUC melalui Stratified 5-Fold Cross-Validation. Interpretasi model dilakukan menggunakan model-based feature importance dan permutation importance.

Hasil penelitian menunjukkan bahwa XGBoost Tuned dipilih sebagai model utama dengan ROC-AUC data uji sebesar 0,864, F1-score kelas corrosive sebesar 0,571, serta rata-rata ROC-AUC validasi silang sebesar 0,904 dengan standar deviasi 0,053. Berdasarkan sintesis EDA, korelasi parameter--target, model-based feature importance, permutation importance, dan interpretasi fisik-kimia, C2H2 dan water content menunjukkan kontribusi prediktif yang paling konsisten dalam membedakan kelas corrosive dan non-corrosive. C2H2 diinterpretasikan sebagai indikator diagnostik dengan daya pembeda yang kuat dan tidak sebagai penyebab langsung pembentukan corrosive sulfur, sedangkan water content memiliki relevansi prediktif dan fisik-kimia serta relatif dapat dikendalikan melalui tindakan pemeliharaan. Breakdown Voltage (BDV) tetap memberikan kontribusi penting sebagai parameter pendukung untuk memverifikasi kondisi dielektrik minyak isolasi. Hasil penelitian ini memberikan dasar analitis bagi pemantauan kondisi transformator dan pengembangan strategi pemeliharaan berbasis kondisi.

Power transformers are critical assets in electric power systems, and their reliability is influenced by the condition of transformer insulating oil. One degradation phenomenon that requires attention is corrosive sulfur, in which reactive sulfur compounds in the oil react with copper materials and may form copper sulfide (Cu2S) deposits. These deposits can reduce the dielectric strength of the oil--paper insulation system and increase the risk of transformer failure. This study aims to identify oil analysis parameters with dominant predictive contributions to distinguishing corrosive and non-corrosive conditions and to analyze their operational roles and controllability.

The research dataset consists of 137 historical observations from six power transformers covering the period from 2012 to 2025 and includes physicochemical oil parameters, Dissolved Gas Analysis (DGA), and Furan (2FAL). The research stages include data selection, binary target encoding, Exploratory Data Analysis (EDA), evaluation of the relationship between numerical parameters and the binary target using point-biserial correlation, and classification modeling using Logistic Regression, Random Forest, and XGBoost. Missing-value imputation and data transformation were performed within the modeling pipeline to prevent data leakage. Model performance was evaluated using test-set ROC-AUC, F1-score with the corrosive class as the positive class, and the mean and standard deviation of ROC-AUC obtained through Stratified 5-Fold Cross-Validation. Model interpretation was conducted using model-based feature importance and permutation importance.

The results show that XGBoost Tuned was selected as the primary model, achieving a test-set ROC-AUC of 0.864, a corrosive-class F1-score of 0.571, and a mean cross-validation ROC-AUC of 0.904 with a standard deviation of 0.053. Based on the synthesis of EDA, parameter--target correlation, model-based feature importance, permutation importance, and physicochemical interpretation, C2H2 and water content demonstrated the most consistent predictive contributions to distinguishing the corrosive and non-corrosive classes. C2H2 is interpreted as a diagnostic indicator with strong discriminative capability rather than as a direct cause of corrosive sulfur formation, whereas water content has both predictive and physicochemical relevance and is relatively controllable through maintenance actions. Breakdown Voltage (BDV) remains an important supporting parameter for verifying the dielectric condition of the insulating oil. These findings provide an analytical basis for transformer condition monitoring and the development of condition-based maintenance strategies.

Kata Kunci : Corrosive Sulfur, Minyak Isolasi Transformator, Oil Analysis, XGboost, Feature Importance.