Case C7 - Fault detection for power transformers
An efficient and working cooling system is important to limit the temperature of the oil in power transformers, since high temperatures cause aging of the winding insulation paper. The aim of this case was to develop and test models for monitoring the performance of the transformer cooling system. To this aid, a feed forward neural network model was developed to predict the top oil temperature from the transformer load and the cooling water temperature. Comparing this prediction with actual values can then identify possible cooling system fault. All data used in the case were from Skagerak's Uvdal 1 power plant. As seen in the figure, the trained model for the transformer top oil temperatures is not that accurate. However, it still follows most trends, and with a suitable threshold value, the accuracy may be good enough to be used to discover severe degradation in cooling performance.
There are multiple possible causes for the inaccuracy. The top oil temperature is only predicted from two parameters, and there may be other factors affecting the temperature. The internal design and temperature sensor placements of transformers varies, and this affects the extent to which the top oil temperature is governed by the transformer load and cooling water temperature. Since the load on the transformer varies a lot, the transformer is in general not in steady state. Hence, a time-dependent model may be more suitable. To this aid, a model was also developed using a recurrent network. This model performed only slightly better than the simple feed forward network. The improvement may have been limited by a time resolution of only one hour that not necessarily captures all important dynamics.
Prediction of the top oil temperature of the transformer in Uvdal 1 power plant using a feed forward network.