Study on Transformer Temperature and Vibration Pattern Analysis Using Regression Algorithms

Recently, Transformers, being critical components in power systems, require accurate condition monitoring to prevent failures. However, conventional diagnostic approaches often rely on periodic inspections or failure analyses, limiting real time responsiveness. This study proposes a regression based condition prediction model using an Autoregressive Neural Network (AR-Net), which effectively captures dynamic patterns in time-series data. The model utilizes composite sensor data specifically temperature and vibration collected under normal operation, enabling non-intrusive diagnostics. Compared to traditional regression and RNN models, AR-Net offers improved adaptability to shifting operational trends. Experimental results show excellent prediction accuracy, with Mean Absolute Percentage Error (MAPE) of 0.93% for temperature and 5.09% for vibration each. These results show the feasibility of applying AR-Net for real time transformer condition prediction and support its integration into intelligent predictive maintenance systems.