Study On Transformer Fault Diagnosis Based On M-LS-SVM

Oil-filled transformer is an important power system equipment, the reliability of its operation is directly related to the security and stability of the power system, and thus how timely and accurate diagnosis of transformer fault has been the focus of scholars at home and abroad. Oil-filled transformer fault types are closely related to the characteristic gases dissolved in transformer oil, so we can analyze the value of the dissolved gas in the oil for transformer fault diagnosis. During the diagnosis, the first thing is to predict the transformer operational status; the second is to judge fault type for the abnormal transformer.Transformer oil dissolved gas analysis (DGA) method can obtain the concentration data of dissolved gas without affecting the normal operation of the transformer. There are two main forms of its realization, off-line artificial test and on-line monitoring device automatically acquisition. On-line monitoring device has many advantages, such as labor-saving, easy automating management and so on, so it is the trend of development. But this kind of device also has its own shortcomings, mainly in the own equipment high failure rate, systematic biases of measuring data among different monitoring devices and so on, this has led to a low diagnostic accuracy when we apply traditional diagnostic methods in on-line monitoring data. Therefore we need an intelligent diagnosis method that can be implemented on on-line monitoring device according to its characteristics and failure characteristics.At present, some literature has used neural networks, support vector machines and other machine learning methods for transformer fault diagnosis, and achieved good results. But these methods mentioned above are shown in local minimum, kernel selection difficult and other problems, leading to deficiencies appear in transformer fault diagnosis. Aiming transformer fault diagnosis, this paper proposed the mixed kernel support vector machine (M-LS-SVM) model, namely using a linear kernel and a nonlinear kernel adaptive combination in LSSVM model, improve the ability to adapt the complex data, while reduce the hassle for kernel choice. Finally, by using the M-LS-SVM for two instances of the judgment of transformer fault types and the concentration prediction of dissolved gas, the effectiveness of this new approach was demonstrated.