Research On Synchronous Condenser Fault Diagnosis Algorithm Based On RBF Neural Network

Nowadays,with ultra-high voltage direct current(UHVDC)transmission projects being built on a large scale in China,the issue of short of dynamic reactive power reserve and voltage support becomes increasingly prominent.Compared to other reactive power compensation method,synchronous condenser(SC)has larger short circuit capacity and can provide larger reactive power.The application of large synchronous condenser in the power grid has been put on the agenda.However,due to complex structure and lots of parts,large synchronous condenser is prone to fault.Once a fault occurs,it will operate abnormally and affect the stability of power system.Therefore,this paper focuses on the fault mechanism and the fault diagnosis algorithm of synchronous condenser.Firstly,through analysis of the air gap flux density,characteristics of stator or rotor vibration and loop current between parallel branches are obtained,when SC operates on inter-turn short circuit fault,of stator or rotor windings,and rotor eccentricity fault.Secondly,compared to Fast-Fourier-Transform and Short-Time-Fourier-Transform,wavelet transformation and wavelet packet transformation perform better time and frequency resolution on non-stationary signals.Moreover,wavelet packet transformation performs better than wavelet transformation.Wavelet packet transformation is applied to extract the feature vector of SC signal.Principal component analysis algorithm(PCA)is applied to reduce dimensionality of feature vector.Thirdly,training samples before and after dimension reduction are used to train RBF neural network.Training results and network output results demonstrate the feasibility of PCA applying on RBF neural network,and point out shortcoming of RBF neural network algorithm based on K-means clustering algorithm.Finally,this paper proposes an improved RBF neural network algorithm based on improved K-means clustering algorithm,in terms of the randomness of initial center selection and the blindness of number of hidden layer neurons selection.The improved algorithm is applied to training samples after dimension reduction.Simulation results prove the feasibility of the improved algorithm.