Study on thermal aging condition assessment of oil-paper insulation based on statistical features of partial discharge

A thermally accelerated aging experiment was designed for simulating the actual environment of transformer insulation in service. Pulse current method was used to measure partial discharge of the test samples with different aging degrees. Four PD spectrums of PRPD model and its 29 features were extracted for aging condition assessment. Aging condition of oil-paper was defined and normalized into a unit circle. Besides, the aging radius R which represents the condition of aged oil-paper was obtained by a linear transformation on degree of polymerization (DP). According to DP, the aging condition of oil-paper was divided into five stages. By using the aging radius R as a single objective, an assessment model was established based on a three-layer BP neural network which was trained by genetic algorithm and LM algorithm respectively. Finally, the test samples were assessed by the trained GA-BP and LM-BP network. Results demonstrate that GA-BP network has the advantages of higher recognition rates and better generation capability comparing with LM-BP network.