Abstract: A top oil temperature prediction method for UHV shunt reactor based on optimal segmentation and improved semi-physical model is proposed to effectively evaluate the internal thermal state of the UHV shunt reactor. Firstly, based on fully studying the "sinusoidal" change trend of the top oil temperature curve of the UHV Shunt reactor, the top oil temperature curve is segmented by the K-means clustering method. Then the validity function is established based on the intra-segment distance, the inter-segment distance and the overlap degree, and the selection of the optimal number of top oil temperature curves is realized. Finally, after comprehensively considering the discretization error of the semi-physical forecasting model of the transformer top oil temperature and the main influencing factors of the UHV shunt reactor, an improved semi-physical model suitable for the prediction of the UHV shunt reactor top oil temperature is proposed, and the Elman neural network is used to forecast the UHV shunt reactor top oil temperature in East China. The results show that the average forecasting error of the method proposed is relatively high with the forecasting accuracy of 1.00%, which can meet the accuracy requirements of the actual application in the field, thus verifying the effectiveness of the method.