Research on Power Transformer Health Indicator Based on Improved Convolutional Neural Network

To achieve accurate prediction of the remaining lifespan of a power transformer, it is essential to develop sophisticated health indicator (HI) that capture the entire lifespan of the power transformer. In this regard, we propose a method for constructing a power transformer HI based on ICNN-DRSN (Improved Convolutional Neural Network and Deep Residual Shrinkage Network). This approach leverages the ICNN and DRSN to extract dynamic features from online dynamic monitoring data, while utilizing a spectral clustering algorithm to cluster static features derived from data such as power transformer operation settings and environmental conditions. Subsequently, we fuse the extracted dynamic features with the clustered static features to derive a comprehensive HI that effectively characterizes the entire lifespan of transformers. These robust HI serve as the foundation for predicting the remaining lifespan of power transformers. Through example analysis and calculations, we demonstrate that by integrating multiple degradation variables, the fused HI provide a more comprehensive and accurate prediction of the remaining lifespan of transformers, resulting in significant improvements in prediction accuracy.