Abstract: Aiming at the testing problem of dissolved gases concentration in transformer oil, a new prediction modeling method based on mutual information(MI)and kernel entropy component analysis(KECA)was proposed. Firstly, normalized mutual information feature selection method was used to select input variables and the phase reconstruction space were reconstructed for them. Then, feature extraction was carried out in the phase reconstruction space by using KECA, meanwhile, the kernel parameter of KECA was determined by Renyi information entropy. At last, kernel entropy components were extracted by KECA and then they were used as the inputs of extreme learning machine(ELM)which was employed to forecast dissolved gases concentration. Experimental results show that compared with grey model, support vector machine(SVM)and BP neural network(BPNN), the proposed model can sufficiently utilize the dissolved gases information, thus it has a better prediction and generalization.

Key words: dissolved gas-in-oil analysis, mutual information, machine learning, Renyi entropy, kernel entropy component analysis, extreme learning machine