Hybrid Random Forest and Long Short-Term Memory based Transformer Fault Diagnostics and Prognostics

• Author(s): Orji, Osita Lawrence ; Ashigwuike Evans ; Ejimofor Chijioke
• Paper ID: 1707497
• Page: 489-510
• Published Date: 15-03-2025
• Published In: Iconic Research And Engineering Journals
• Publisher: IRE Journals
• e-ISSN: 2456-8880
• Volume/Issue: Volume 8 Issue 9 March-2025

Abstract

This research presents the development and evaluation of a hybrid approach combining Random Forests (RF) and Long Short-Term Memory (LSTM) neural networks for transformer fault diagnostics and prognostics. The study aimed to leverage RF's capability in handling complex, high-dimensional data and LSTM's strength in capturing temporal dependencies to create a more robust and accurate system. The methodology involved comprehensive exploratory data analysis, individual implementation of RF and LSTM models, and their integration through both weighted average and stacked ensemble approaches. The results demonstrated that while the RF model achieved perfect classification accuracy (1.0000) and exceptional prognostic performance (R² = 0.9996, RMSE = 0.0189), the LSTM model showed strong classification capabilities (0.9656) but struggled with remaining useful life (RUL) prediction (R² = -30.0272). The final stacked ensemble hybrid model successfully combined the strengths of both approaches, achieving a fault classification accuracy of 0.940909 and remarkable RUL prediction metrics (R² = 0.999633, MAE = 4.807955 days). This performance significantly surpassed existing approaches in the field, including fuzzy-neural networks and PSO-optimised LSTM networks. The hybrid model demonstrated particular effectiveness in handling the unique challenges of the Nigerian power infrastructure, offering superior performance compared to previous models specifically developed for this context. The research concludes that the stacked ensemble architecture provides a comprehensive solution for transformer health management, successfully balancing high accuracy, computational efficiency, and interpretability. These findings have significant implications for practical maintenance planning and decision-making in power transformer operations.

Keywords

Random Forest (RF), Long Short Time Memory (LSTM), Artificial Neural Network (ANN), Direct Current (DC), Remaining Useful Life (RUL), Particle Swam Optimization (PSO)

Citations

IRE Journals:
Orji, Osita Lawrence , Ashigwuike Evans , Ejimofor Chijioke "Hybrid Random Forest and Long Short-Term Memory based Transformer Fault Diagnostics and Prognostics" Iconic Research And Engineering Journals Volume 8 Issue 9 2025 Page 489-510

IEEE:
Orji, Osita Lawrence , Ashigwuike Evans , Ejimofor Chijioke "Hybrid Random Forest and Long Short-Term Memory based Transformer Fault Diagnostics and Prognostics" Iconic Research And Engineering Journals, 8(9)