Artificial neural network modelling of partial discharge parameters for transformer oil diagnosis

On-line partial discharge (PD) detection technique is gaining importance as a condition-monitoring test for transformers. Presently on-line condition-monitoring efforts of transformers are primarily directed towards chemical analysis of the oil. In this paper, a new approach has been proposed to interpret the condition of the transformer oil from PD results. An experiment was carried out on a simulated transformer tank to obtain PD data under three different oil conditions (e.g. clean oil, oil with solid contaminants and oil with high moisture content). Besides the oil condition, the loading of the transformer was also simulated by varying the temperature of the oil. The PD parameters obtained from the experiment are PD magnitude (q) and number of counts (n). The experimental results were then modelled using the multi-layer feedforward neural network (NN) with back propagation technique. Once trained, the NN model (i.e. "oil condition = f(q, n)") is then capable of predicting the condition of the transformer oil, under any given set of q, n with a mean absolute error (MAE) less than 5%. The outcome of this research provides an alternative method of diagnosing the condition of the transformer oil without the need for conventional chemical analysis.