Determination of Relative Thermal Performance of Power Transformers Using Data Driven Thermal Models

Reliability of power transformers is essential for managing electrical power transmission systems. Condition monitoring is needed to ensure that power transformers work as expected. A critical parameter used in condition monitoring is the hot spot temperature, which is governed by the thermal performance of the unit. Transformers are type tested during commissioning to guarantee that the temperature rise above ambient conditions does not exceed 78°C at their maximum capacity. However, thermal performance could change over a long period of time due to ageing. Generally, a transformer has sister units that are ordered at the same time and have the same specification, this group of units is referred to as a transformer family. Theoretically, thermal performance of transformers in the same family should be identical. However, due to ageing and different loading conditions, thermal performance could gradually diverge within a family. In this paper, thermal models for power transformers developed using Gaussian process regression are proposed. Once transformer thermal models for each transformer have been built, they are used to predict the transformer temperature of sister units. Typically, the average errors between the measurement and prediction made by the thermal model using its own data should be about zero, however, the average errors that are generated by the thermal model of the sister transformer could be positive or negative, if the thermal behavior between them is significantly divergent. The proposed method has been validated using data for three 400kV/275kV autotransformers and has been shown to work effectively.