1.China Three Gorges University;2.Xi’an Jiaotong University
Accurately identifying the type of transformer fault based on the characteristics of dissolved gas content in the oil is of great significance to its safe operation. Deep Belief Network (DBN) has unique advantages in feature extraction from sample data. In this paper, DBN is used as a fault identification model. The original data set of dissolved gas in transformer oil is directly input to the training model, and the DBN is analyzed through three intelligent search algorithms. The three important parameters of mid-batch processing, gradient descent learning rate, and number of neural units are intelligently optimized to solve the problem of low fault recognition rate when a small amount of raw sample data is directly input. The results show that when using the Cuckoo Search Algorithm (CS) to optimize the DBN model, the effect is better than the commonly used particle swarm search algorithm (PSO) and genetic algorithm search (GA) optimization. The fault total recognition rate of CS-DBN is 4.2% higher than that of GA-DBN, 2.5% higher than PSO-DBN and 56.2% higher in evolution efficiency; the generalization performance of CS-DBN model is also better than the other two.