Improved PSO-BP Neural Network For Power Transformer Fault Diagnosis

As one of the most important power transmission equipments at power system, power transformer's performance directly affects the security and reliability of operation of power system. Regarding to the reliability and economy of power transformer, it is significant to accurately obtain the operation status and fault conditions, and to take appropriate measures to deal with the fault in a timely manner which can improve power system's operation for safety. Though many solutions have been applied to the diagnosis of transformers'failures, they still have many flaws. None of them can completely make the correct diagnosis in short time.In order to solve the existence of major problems which the current Transformer fault diagnosis technology has in practical applications, this paper introduces the neural network theory. The paper combines the improved Particle Swarm Optimization (PSO) algorithm and BP neural network, and applies the algorithm to the Dissolved Gas Analysis (DGA) of transformer fault diagnosis.For the weight of BP neural network training in non-linear, complex process, standard Particle Swarm Optimization algorithm which makes the inertia weight reduction linearly often fails to reflect the actual optimized search process. Dynamic Particle Swarm Optimization algorithm can be used to achieve the non-linear search, but it is easy to fail into local optimization.This paper introduces the particle variation on the algorithm to improve the PSO algorithm, and proposes the Mutational Dynamic Particle Swarm Optimization algorithm. During the process of convergence, when one particle is selected as the global best continuously, the entire swarm will converge to that point which will cause the algorithm into a local optimum. In order to solve this problem of PSO, the Mutational Dynamic Particle Swarm Optimization tracks the global optimal point during the iteration process. Before every iteration, the algorithm checks whether the global optimal point changes. If not, the algorithm will select some particles randomly and make their weights grow at a exponential rate which was decided by the iteration times of a certain particle's being selected as the global optimal point, which make the particle have a velocity to jump local optimization and to global optimization.At last, this paper applies the improved PSO-BP neural network into Transformer Fault Diagnosis. The paper designs a complete Experimental blueprint, the results of which prove the algorithm which this paper proposes has a better performance than other algorithm in Transformer Fault Diagnosis.