A Comparative Analysis of CNN and LSTM Models for Transformer Fault Diagnosis using Dissolved Gas Analysis

The diagnosis of transformer faults through the Dissolved Gas Analysis (DGA) technique has been critical to the safe and reliable use of power systems. In the current research, the comparative analysis of two approaches of deep learning (Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks) is conducted in order to classify transformer faults based on the DGA data. The research will be based on six canonical transformer fault indicators, namely, Partial Discharge (PD), Discharges at different energy levels (D1, D2), and Thermal faults at different temperature levels (T1, T2, T3). The information was normalized and preprocessed and was categorized into classes with regard to training and testing the models. All models were trained in the same conditions with the Adam optimizer. The CNN generated a test accuracy of 92.78 %, and the LSTM achieved a greater accuracy of 97.22 %. A confusion matrix implies that LSTM outperformed CNN in most fault types very consistently. These results show that LSTM has better fault-classification abilities than other classifications when it is used on DGA datasets, and thus, it can be a more efficient foundation of intelligent transformer condition monitoring and predictive maintenance systems. Unlike traditional machine learning or rule-based methods, this work emphasizes a comparative analysis between CNN and LSTM models for DGA-based transformer fault classification, where LSTM proved to better capture temporal dependencies, improving diagnostic accuracy.