Laser Raman spectroscopy is an effective method for detecting the aging state of transformer oil-paper insulation. With the expansion of sample quantity and the gradual increase of data set dimension, it is of great significance to study the evaluation method of oil-paper insulation aging in transformer suitable for high-dimensional Raman spectral data. An oil-paper insulation environment similar to the internal insulation structure of the field transformer was designed, and the accelerated thermal aging experiment was carried out and regularly sampled to obtain ten types of oil samples with increasing aging degrees, then these samples were detected using laser Raman spectroscopy. The compound sparse derivative modeling method was used to preprocess the original Raman spectral data, which can complete the noise elimination and baseline correction in one step. The differential feature selection method was introduced to screen the spectral features with significant changes under different aging degrees, and the variance of the feature point data set with different aging degrees was calculated under the same Raman shift. Furthermore, the Raman feature variable corresponding to the data sequence with a large difference was selected, and the variance threshold was set to 0.5 for feature selection, each sample selected 304 from 1 023 spectral feature points for subsequent analysis. In this paper, many different types of algorithms were introduced to process the high-dimensional sample data set of transformer oil-paper insulation aging Raman spectra. For instance, the K-means clustering algorithm, the Fisher algorithm and Random Forest algorithm were used to establish a model with the preprocessed data of the obtained samples. The evaluation accuracy, lifting degree and Kappa coefficient were introduced to evaluate the discriminant effect of each mathematical model. The results show that supervised learning Fisher algorithm and Random Forest algorithm have a better effect and discriminatory advantage compared with the unsupervised learning k-means clustering algorithm because the discrimination ability of the model is improved by 1.166 6 and 1.95, respectively; Judging from the discrimination accuracy and Kappa coefficient, the discriminant model established by the strong classifier Random Forest algorithm is better than the Fisher discriminant model, for its accuracy is improved by 10%, and the Kappa coefficient is increased by 0.111 5. Compared with a single classifier, a strong classifier composed of multiple single classifiers has better generalization evaluating of transformer oil-paper insulation aging, and the model is more stable and reliable. By comparing three different types of algorithms, the discrimination advantages of the supervised learning strong classifier Random Forest algorithm in evaluating transformer oil-paper insulation aging are determined, which lays the foundation for the effective evaluation of transformer oil-paper insulation aging.