Transformer Faults Multiclassification using Regularized Neural Network and Dissolved Gas Analysis

In electrical power systems, power transformers are one of the most essential equipment required to transmit and distribute electricity from sources to consumers. To ensure the reliability and stability of a power system, the condition of power transformers must be carefully monitored using methods such as the conventional Dissolved Gas Analysis (DGA) interpretation methods. However, various studies have indicated the limitations and challenges of the conventional methods associated with the accuracy and uncertainty of the results due to inconsistent results obtained when using different methods i.e. the Duval Triangle, IEC, CIGRE, Doernenburg, Key Gas, Roger's Ratio, and Nomograph methods. Over the years, there have been developments in integrating machine learning techniques in attempts to address the limitations presented by conventional methods. This study proposes a Regularized Neural Network model that is used to classify transformer faults based on gas concentrations. The MATLAB Workspace interface was used for modeling and analysis for this study. Initially, a standard neural network algorithm is used to classify the faults but it exhibits some substantial misclassification, a regularized neural network is then introduced to address the misclassifications and it correctly classifies all the transformer faults.