Dynamic Fault Prediction of Power Transformers Based on Lasso Regression and Change Point Detection by Dissolved Gas Analysis

Practical features of dissolved gases analysis (DGA) are selected and proposed from 62 key gases combinations through maximal information coefficient (MIC) to minimize the influences of random errors and relative percentages variation for field application. Then the Pearson correlation coefficient is employed to filter and optimize the feature set to reduce the redundancy of the selected features. Lasso regression is proposed to build a multi-dimension linear model of the selected features. In the multi-dimension model, the position in which the parameter changes drastically is defined as a change point, which contains specific information on the transformer's operation status. The case analysis demonstrates that the variation of selected features under abnormal status can be figured out from that of normal status prior to fault occurrence. The change point detection based on Lasso regression shows the least number of days between change point and time of failure and standard deviation (SD), which accurately reflects the location of the transformer fault in most scenarios. Therefore, the proposed technique provides an available approach for the dynamic fault prediction based on the dissolved gas data, showing the advantage of robustness, data-free training, and early warning.