Acceptability of Four Transformer Top-Oil Thermal Models—Part I: Defining Metrics

Eventually, the prediction of transformer thermal performance for dynamic loading will be made using models distilled from measured data, rather than models derived from transformer heat-run tests. Which model(s) will be used for this purpose remains unclear. In this paper, we introduce metrics for measuring the acceptability of transformer thermal models. For a model to be acceptable, it must have the qualities of adequacy, accuracy, and consistency. We assess model adequacy using the metrics: prediction R
²
and plot of residuals against fitted values. To assess model consistency, we use the variance inflation factor (which measures multicollinearity), condition number, eigenstructure, parameter sensitivity, and the standard deviation of model parameters and predicted maximum steady-state load (SSL
_max
). To assess model accuracy, we use the comparison of model parameters with nominal values and error duration curves. Other metrics of interest are also introduced. In a companion paper, these metrics are applied to the four models defined in this paper and a relative ranking of the acceptability of these models is presented.