Prediction of Transformer Top Oil Temperature Based on Kernel Extreme Learning Machine Error Prediction and Correction

For accurate estimation of transformer hot spot temperature (HST) which can provide the essential information and support of loading capacity estimation, thermal fault prevention, and insulation life prediction, we established a model for predicting the transformer top oil temperature (TOT) by the error prediction and correction via kernel extreme learning machine (KELM). KELM was used to predict the TOT prediction error of typical Susa thermal circuit model, thent he prediction value was used to correct the TOT prediction result. To improve the prediction accuracy, the gravitational search algorithm(GS) was used to optimize the punishment coefficient and kernel parameter of KELM. The case study shows that the prediction results of the model proposed in this paper are in accordance with the measured results and are superior to results of semi-physical model without error correction- Susa thermal circuit model, and are also superior to results of model directly based on nonlinear fitting and regression method-KELM optimized by GS. The training time of KELM optimized by GS is substantially less than support vector machine optimized by GS and Elman neural network, and the prediction accuracy is slightly superior. © 2017, High Voltage Engineering Editorial Department of CEPRI. All right reserved.