Insulation fault diagnosis of power transformers by genetic programming based polynomial networks model

In this paper, a Genetic Programming based Polynomial Networks Model (GPPNM) is proposed to promote the diagnostic performance of incipient insulation fault of power transformers using dissolved gas analysis (DGA). The proposed GPPNM constructs the polynomial networks by tree-like structure of Genetic Programming (GP) instead of conventional hierarchical architecture. Each node in the GPPNM is flexibly selected from a set of second-order polynomial functions and feature variables. The structure of GPPNM evolves by generations in the global search space according to its fitness value to learn the complex and numeric relationships between dissolved gases and fault types. Based on the analysis of insulation fault properties, a hierarchical classification strategy is built to diagnose fault types step by step. In comparison with results of conventional IEC method, artificial neural networks method and self-organizing polynomial networks (SOPN) method, the proposed GPPNM can automatically generate optimal model structure and shows better performance.