Power Transformer Fault Diagnosis from Dissolved Gas Analysis Using Machine Learning

Faults inside oil-filled power transformers - partial discharge (PD), discharges of low (D1) and high energy (D2), thermal faults below 300 C (T1), 300-700 C (T2) and above 700 C (T3) - produce characteristic gases (H2, CH4, C2H6, C2H4, C2H2) that dissolve in the oil. This project diagnoses the fault type from those five gases and compares the classical interpretation methods (Duval triangle 1 with the full zone polygons, Duval pentagon 1, IEC 60599 three-ratio method, Rogers ratios) with SVM, Random Forest, LightGBM, k-NN and logistic regression. THE DATA ARE REAL PUBLISHED CASES whose fault was found by inspection: 1,529 de-duplicated cases from 7 sources (IEC TC 10 database, Duval 2002 tables, IEEE DataPort, NCEPR case book, NE Grid and Fujian utility records, Nanfak et al. 2024), main set 1,199 cases with abnormal gas levels. In 10 x 5-fold cross-validation grouped by transformer, the deployed Random Forest reaches 71.0 +/- 4.5 % against 52.6 +/- 4.4 % for the Duval triangle; on the IEC TC 10 cases the Duval triangle reproduces the literature (86.8 %). Leave-one-source-out shows that the ML gain does not transfer to a new source. A GNU Octave cross-check (run, 100 % identical), a MATLAB version (written, not run), an Excel / LibreOffice workbook and a web application with five databases are included. Everything runs on a student laptop CPU.

Category: Electrical and Electronics Engineering (High Voltage Engineering, Power Transformers, Dissolved Gas Analysis, IEC 60599, Duval Triangle, Duval Pentagon, Rogers Ratios, IEEE C57.104, Machine Learning, SVM, Random Forest, LightGBM, GNU Octave, MATLAB, Python)

Electrical & Electronics Engineering Project EE10 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.