Transformer linear thermal modeling

Improving the utilization of transformers requires that the transformer top-oil an hot-spot temperatures (TOTS and HSTs) be predicted accurately. Traditionally, temperature prediction used the ANSI/IEEE Clause 7 model, which relied on the transformer test report. This model has two primary shortcomings: it relies on measurements that may be inaccurate; and it does not accurately model the ambient temperature dynamics. This work discusses some challenges encountered when building more accurate linear models from field measured data.; This research work centers around three primary thrusts: defining a metric for measuring model reliability, developing a means for calculating model reliability and exploring methods for improving model reliability.; Steady-state predicted load is shown to be an effective metric for measuring model reliability. It correlates better than other proposed metrics with maximum dynamic predicted load, which is the ultimate goal of building transformer dynamic thermal models.; It is shown that the classical means of defining model reliability from the least squares regression fails for transformer dynamic thermal models built from measured data and another sample-based method is developed.; Much of the research effort focused on improving model reliability through various means, including: high break-down estimators, Lp estimators, the Cochrane-Orcutt regression, including solar radiation, transformer taps, and temperature dependence in the model, filtering of input data to remove high frequency noise, data quality control, and model screening. Of these efforts, the most successful were data quality control, data-set screening, and the Cochrane-Orcutt regression.; Data quality control and data-set screening are shown to improve reliability significantly. They improve the reliability of the steady state load prediction by 27%.; The ordinary least squares method is biased and inconsistent when applied to problems like ours, which have lagged dependent variables and autocorrelated noise. The Cochrane-Orcutt regression provides an unbiased and consistent estimator when applied to this problem and is shown to be effective when applied to simulated data with large amounts of autocorrelated noise. The field data we used in this work is of good quality, so the noise proved not to be prohibitively large, but is found to be highly correlated.