Study On Fault Diagnosis Approaches For Oil-immersed Power Transformers

Power transformers are among the key equipment in electrical power transmission/distribution systems, and to detect early incipient faults in transformers timely and accurately is of great significance for enabling reliable operations of power systems. On the basis of the analysis of the fault mechanism and the existing fault diagnosis methods, self-organization antibody net (soAbNet), extreme learning machine (ELM) and evidence theory has been studied and applied to fault diagnosis of power transformer in this dissertation. The main content of the thesis are as follows.There is no network compression mechanism in the learning algorithm of soAbNet. As a result, there are many redundant antibodies in the trained immune network. To solve this problem, an improved soAbNet was proposed. In this method, a network compression mechanism was introduced into soAbNet, and the distribution of memory antibodies in immune network was adjusted with an affinity threshold. Experiment results show that the network compression mechanism could accelerate the convergence speed of soAbNet, optimize the network structure and maintain the stability of the immune network simultaneously.Because of the initial antibodies are randomly selected in soAbNet, its network performance is instability when soAbNet is applied to transformer fault diagnosis. A hybrid immune network was proposed for transformer fault diagnosis, which is a combination of soAbNet and immune operator. Immune operator obtains initial antibodies using K-means optimal clustering algorithm, and the affinity threshold is optimized by using particle swarm optimization (PSO) algorithm. Experiment results show that the performance of hybrid immune network is more stable than that of soAbNet, and the diagnostic accuracy is higher than that derived from individual diagnosis approaches.However when the proposed hybrid immune network is applied to transformer fault diagnosis, it requires a large number of calculations to obtain initial antibodies, its learning speed is slow, and it is difficult to determine its learning parameters. Therefore, ELM and kernel-based ELM (KELM) were applied to transformer fault diagnosis, and the learning parameters of KELM were optimized by using PSO. Experiment results show that the diagnostic accuracy of ELM is a bit higher than that derived from hybrid immune network, and its learning time is far less than the latter, unfortunately its performance is instability; compared with other intelligent diagnosis approaches, KELM requires the least training time and testing time, its correct diagnosis rates on training dataset and testing dataset are the highest, and its performance is very stable.Since the relations between characteristic gases and transformer fault are complicated and fuzzy, the results of different diagnosis techniques might be inconsistent. A fault diagnosis for power transformer based on evidence theory was proposed to combine the diagnosis results of hybrid immune network, KELM and fuzzy theory by using combination rule for conflicting evidence, and it could effectively integrate multiple diagnosis techniques with multiple feature information. Experiment results show that the proposed fusion method could improve the reliability of transformer fault diagnosis.