When a transformer's windings get too hot, either load has to be reduced (i
n the short term) or another transformer bay needs to be installed (in the
long run). To be able to predict when either of these remedial schemes must
be used, we need to be able to predict the transformer's temperature accur
ately, Our experimentation with various discretization, schemes and models,
convinced us that the linear and nonlinear semiphysical models we were usi
ng to predict transformer temperature were near optimal and that other sour
ces of input-data error were frustrating our attempts to reduce the predict
ion error further, In this paper we explore some of the sources of error th
at affect top-oil temperature prediction. We show that the traditional top-
oil rise model has incorrect dynamic behavior and show that another model p
roposed corrects this problem. We show that the input error caused by datab
ase quantization, remote ambient temperature monitoring and low sampling ra
te account for about 2/3 of the error experienced with field data. It is th
e opinion of the authors that most of this difference is due to the absence
of significant driving variables, rather than the approximation used in co
nstructing a linear semiphysical model.