A New Evolutionary Neural Network Algorithm Based on Improved Genetic Algorithm and its Application in Power Transformer Fault Diagnosis

In order to overcome the limitation such as premature convergence and low global convergence speed etc. of genetic algorithm, some improvements were made for classical genetic algorithm. The genetic individuals were separated into male individuals and female individuals, and consanguinity was fused into individuals. Two individuals with different sex could reproduce the next generation only if they were distant consanguinity individuals. A help operator was used to help individuals according to the given probability. Based on the improved genetic algorithm, a thoroughly evolutionary neural network algorithm named IGA-BP algorithm was proposed. In IGA-BP algorithm, genetic algorithm was used firstly to evolve and design the structure and all training parameters of neural network roundly. Then, training samples were used to search the optimal solution again. The disadvantage of neural network that its structure and training parameters were decided stochastically or by one´s experience was overcome. IGA-BP algorithm was used to diagnose power transformer faults. A fault diagnosis model of power transformer was founded based on neural network. The illustrational results show that this algorithm is better than traditional BP algorithm in both speed and precision of convergence. We can realize a fast and accurate diagnosis for power transformer faults by this algorithm.