Abstract:For the situation that most of the data of transformers with low running time are normal data, in order to utilize the historical normal state data to identify efficiently whether the new data is abnormal, a power transformer abnormal state recognition model based on improved K-means clustering, is proposed in this paper. Now that most of the power transformer operation data are normal state data, and the normal state data gradually changes according to a certain trend while the abnormal state data changes rapidly, based on the historical normal data and K-means clustering, a recognition model for power transformer is established. According to the clustering results of normal data, thresholds can be acquired to identify new data .By calculating the distances from the new data to the cluster centers and comparing the distances with the thresholds to determine whether the transformer is abnormal. To solve the problem of traditional K-means algorithm, an improvement about choosing K value and initial cluster center based on data density and distance is proposed, which can give K-means clustering stable cluster centers and K value to make the procession of clustering quicker, more stable and efficient. The example analysis shows that the recognition model can effectively identify the abnormal state of the power transformer quickly and accurately and that the model provides a new idea for the state assessment of transformers.