Shunt reactor is an important power equipment in the power grid. Accurate estimation of its oil temperature variation trend can provide an important basis for early fault monitoring and warning. An kind of oil temperature prediction model of shunt reactor based on feedback-constrained variational mode decomposition (VMD) and long short-term memory neural network (LSTM) is proposed. Firstly

the signal feedback constrains and weighted sample entropy are used to optimize VMD decomposition number k and penalty factors α to form a VMDFS decomposition method

decompose the original oil temperature sequence of the shunt reactor into multiple sets of smooth sub-sequence so to eliminate the influence of the unsteady information. Then

the LSTM neural network prediction models for each sub-sequence is set up and particle swarm is used to optimize the number of neurons. Finally

the predicted oil temperature of each sub-sequence is superimposed to obtain the final estimated oil temperature of the shunt reactor. It is proved through the analysis of the measurement data of a 220 kV shunt reactor at a substation that the proposed model has such advantages as high accuracy

strong robustness and slow accumulation of errors in single-step and multi-step estimation performance

and can effectively reflect to the oil temperature variation trend of shunt reactor

as well as provide reference for its monitoring and early warning.