Transformer fault diagnosis research based on LIF technology and IAO optimization of LightGBM

摘要

Transformer fault diagnosis is a necessary operation to ensure the stable operation of a power system. In view of the problems of the low diagnostic rate and long time needed in traditional methods, such as the dissolved gas in oil method, a laser-induced fluorescence (LIF) spectral technology is proposed in this paper, which incorporated an improved aquila optimizer (IAO) and light gradient boosting machine (LightGBM), to predict the types of transformer faults. The original AO was improved using the Nelder Mead (NM) simple search method and opposition-based learning (OBL) mechanism, which could improve the parameter optimization ability of the model. Normal oil, thermal fault oil, local moisture oil, and electrical fault oil were selected as experimental samples. First, the spectral images of the four oil samples were obtained by LIF technology, and the fluorescence spectral curves obtained were preprocessed by multivariate scattering correction (MSC) and normalization (normalize), while kernel-based principle component analysis (KPCA) was used for dimensional reduction. The dimensionality-reduced data were then imported into the LightGBM model for training, and the IAO algorithm was used to optimize the parameters of the LightGBM. Finally, the experiment showed that the LIF technology demonstrated good recognition of the fault types for transformer fault diagnosis; the data purity after MSC preprocessing was higher than that of other processing methods; the prediction effect of the LightGBM model was superior to other prediction models; the LightGBM model optimized by IAO had better convergence, parameter optimization ability, and prediction accuracy than the LightGBM model optimized by the original AO and particle swarm optimization (PSO). Among the models, the MSC-IAO-LightGBM model had the best effect on fault prediction, with the mean square error (MSE) reaching 9.0643 x 10(-7), mean absolute error (MAE) reaching 8.7439 x 10(-4), and goodness of fit (R-2) approaching 1. It can be implemented as a new diagnostic method in transformer fault detection, which is of great significance to ensure the stable and safe operation of power systems.