Investigating the Remaining Useful Life of Power Transformer Using Random Forest and Long Short Term Memory

Abstract

Operational factors such as load imbalance, hot-spot temperature, dynamic derating, and voltage loss have been shown to accelerate insulation degradation, which has a direct impact on the reduction of the Remaining Useful Life (RUL) of the transformer. However, conventional RUL prediction approaches are generally still static and have not been able to represent actual operational conditions that are dynamic and change over time. This study aims to develop a transformer RUL prediction model based on Dynamic Failure Mode and Effects Analysis (Dynamic FMEA) with the integration of machine learning methods. This approach utilizes Mean Decrease in Accuracy (MDA) to measure the relative importance of each operational parameter to the risk of failure, resulting in a dynamic Risk Priority Number (RPN) that is adaptive to the actual conditions of the transformer. The hybrid model is built by combining Random Forest as a base model for degradation pattern extraction and Long Short-Term Memory (LSTM) to capture the temporal dynamics of RPNs based on historical operational data. The data used in this study include phase current, load imbalance, hot-spot temperature, dynamic derating, and voltage loss (?V%). The resulting dynamic RPN values are then mapped into an exponential model to estimate condition-based RULs. The results show that the integration of Dynamic FMEA, MDA, and machine learning models can produce RUL estimates that are more adaptive and informative than the static approach, and have the potential to support the implementation of risk-informed maintenance to improve the reliability and sustainability of electric power systems.