Application of local memory-based techniques for power transformer thermaloverload protection

Power transformers are some of the most expensive components of electrical power plant. The failure of a transformer is a matter of significant concer n for electrical utilities, not only for the consequent severe economic los ses but also because the utility response to a customer during the outage c ondition is one of the major factors in determining the overall customer at titude towards the utility. Therefore. it is essential to predict the therm al behaviour of a transformer juring load cycling and in particular in the presence of overload conditions. The authors propose a novel technique to p redict the winding hottest spot temperature of a power transformer in the p resence of overload conditions, as an alternative methodology to the radial basis function network (RBFN) based technique presented in a precious pape r. The method proposed is based on a modified local memory-based algorithm which, working on the load current, the top oil temperature rise over ambie nt temperature and caking into account other meteorological parameters, per mits the recognition of the hot spot temperature pattern. In particular som e corrective actions for the classical local methods M;ill be evidenced to customise it for real-time applications. Data obtained from e experimental tests allow the local learning algorithm to be tested to evaluate the perfo rmance of the proposed method in terms of accuracy.