Abstract: In order to realize the accurate early warning and intelligent monitoring of transformer oil insulation state, by analyzing the dissolved gas analysis data in transformer oil sent for inspection by some power plants in Inner Mongolia over the years, a transformer oil insulation state prediction method based on feature evaluation and improved SOA(Seagull Optimization Algorithm) optimized DELM(Deep Extreme Learning Machine) model is proposed, so as to accurately predict the dissolved hydrogen and total hydrocarbon content in transformer oil during operation. In terms of feature extraction, the correlation degree between the features is evaluated by calculating the mutual information between the input vector and the predicted output, and the feature with the highest correlation degree constitutes the simplest input vector. In terms of predicted output, by adding additional variable and improving the selection of the SOA parameters, the algorithm can converge quickly and avoid falling into local optimal, so as to realize the network weight and hidden layer bias optimization of DELM prediction model. Finally, compared with other prediction models, the historical samples of the seven power plants are analyzed in sequence, and the applicability of the proposed method is verified comprehensively.