Prediction of top oil temperature for transformers

When a transformer's winding gets too hot, either the load has to be reduced as a short-term solution or another transformer bay needs to be installed as a long-term plan. To decide on whether to deploy either of these two strategies, one should be able to predict the transformer temperature accurately. In this work, the traditional top-oil-rise model, top-oil model (which includes an ambient temperature) and semi-physical top-oil model are compared. The semi-physical top-oil model outperforms the other two models. Several attempts are also reported to improve the model used for top-oil temperature (TOT) prediction. It is shown that regardless of the order or complexity of the model, no model performs significantly better than the semi-physical top-oil model investigated; moreover, many models have performance measures that are approximately the same as the semi-physical model.;Some of the sources of error that affect top-oil temperature prediction are studied here. Experimentation with various discretization schemes and models convinces the author that the semi-physical top-oil model used to predict transformer temperature is near optimal and that other sources of input-data error are frustrating the author's attempt to reduce the prediction error Further. The research demonstrates that the input error caused by database quantization, remote ambient temperature monitoring and low sampling rate accounts for about two-thirds of the error experienced with field data. The results of these simulations also show that the error caused by these sources is less than that obtained when using equivalent field data. It is the opinion of the author that most of this difference is due to the absence of significant driving variables, rather than the approximation used in constructing a semi-physical model.;To further improve the error performance of the semi-physical top-oil model, three different neural network models including static neural network, temporal processing network and recurrent neural network models are examined for TOT prediction. Of the three neural network models, the recurrent network provides the best performance consistently in terms of both the mean-square error (MSE) and the peak error. The performance of both the recurrent neural network model and the semi-physical top-oil model are comparable. The preferred model for predicting TOT is the linear semi-physical model because it permits the use of simple and robust linear regression techniques.