In order to estimate the thermal condition of transformer accurately to guide loading operation

we put forward a mode for the top oil temperature（TOT） of transformers based on the hybrid kernel extreme learning machine（HKELM）. The Kernel extreme learning machine is used to achieve fitting and regression analysis of the mapping relation between TOT and its impact factors. The model takes main influential factors of the TOT including ambient wind velocity into account. Moreover

we adopt hybrid kernel parameters to let the learning machine have better learning ability and generalization performance for improved model precision. The particle swarm optimization（PSO） is adopted to train the model and optimize the hybrid kernel parameters

and the positive and negative errors of the predicted values are processed with different allowable limits so that predicted values are greater than measured values. Therefore

the precision is improved and the predicted results are much more conservative and reliable. The case studying for measured data of different seasons suggests that the prediction results of the model proposed with positive error all the time are in accordance with the measured results. The maximum error of the model is 1.97 centigrade degree

is 78.49% of the BP neural networks model and 82.43% of the least square support vector machine model in the same condition

respectively. The model possesses better prediction accuracy of the TOT

achieving the thermal condition estimation of the power transformer more accurately and appropriately.