Evaluation on residual breakdown voltage of generator insulation by RBFNN

A radial basis faction neural network (RBFNN) is built to evaluate the risidual breakdown voltage of generator stator insulation. The input parameters of RBFNN are selected from the nondestructive parameters, which have big related coefficient. To investigate the nondestructive insulation diagnosic data, the accelerated multi-stress aging experiments have been carried out on some actual stator bars of generator. AC breakdown voltages of the bars are detected after the nondestructive data has been measured. According to the selection rule, Δtanδ, ΔC, S_k+ and S_k- of Hqn (ψ) are selected as the input parameters of RBFNN, and the output parameter is the residual breakdown voltage. The nondestructive parameters from twenty-four samples of bars have been used to train, exam the developed RBFNN and to predict the residual breakdown voltage by it. These samples are selected randomly, thus these parameters should be standardized to reduce their dispersity before being used. The numerical results show that the biggest relative error between the prediction by RBFNN and the measured breakdown voltage is smaller than 6% and the average relative error is smaller than 3%. The residual breakdown voltage is useful for estimating the residual life of large generator insulation. According to the IEC standard, the life of stator insulation arrives at the end point if the residual breakdown voltage falls down to 50% of its initial value. This research may be important to estimate condition of the insulation and to predicate residual life of stator insulation in the case of a few available samples.