Improving the utilization of transformers requires that the hot-spot temepe
ratures and top-oil temperatures (TOTs) be predicted accurately. Our experi
mentation with various discretization schemes and models proved that many o
f the linear and nonlinear semiphysical and nonphysical models we were usin
g to predict transformer TOT were correctly modeling the TOT behavior. Our
experience convinced us that noisy input data and the absence of data on si
gnificant driving variables, not model deficiencies, were frustrating our a
ttempts to reduce the prediction error further. In this paper, we discuss t
he body of research that leads us to these conclusions.