Low-Frequency Oscillations Model Analysis Based On Neural Network Filter

Low frequency oscillation problem induced by interconnection of large powergrids is becoming more and more serious, which has seriously affected the safeoperation of power grid. Since the actual system parameters are very complicated, theimportant premise to analyze mechanism of the low frequency oscillation and thesuppression is to identify low frequency oscillation mode by actual measured data ofthe running state. The commonly used Prony method is sensitive to noise. Therefore,this article mainly aims at the research of filtering pretreatment of the low frequencysignal with noise.This article introduces the basic principle of Prony method and mathematicalderivation of the model and analyzes the parameter selection of sampling frequency,the length and effective order. Simulation examples are analyzed in the background ofno noise, white noise, stochastic pulse noise and high frequency harmonic noise. Itshowed that Prony method is a effective method for analysis of low frequencyoscillation, but it is more sensitive to noise and the error is large when identify thesignal with noise.This article proposed the cosine basis neural network filtering method which isbased on the analysis of bandwidth. First of all, approximate the low frequencyoscillation signal through the cosine basis neural network and it can determine thesignal bandwidth range through the analysis of weight. Then band-pass filteraccording to the signal bandwidth, and import the output signal into Prony module.Aiming at the determination of effective bandwidth, this article presents methods withthe fixed bandwidth and dynamic bandwidth. The simulation examples are analyzed inthe background of impulsive noise, high frequency harmonic noise and random whitenoise. It showed that the method has a better filtering effect, and it has positivesignificance to improve the accuracy of pattern recognition of the Prony method.This article proposed the FIR adaptive filtering method which is based on theneural network. According to the FIR filtering theory, make the wide areameasurement signal as input and the past and the current input values of the signalwith noise as hidden layer neurons. Using the recursive least squares method to trainthe weights, so as to achieve the effect of filtering. In addition, regulating theperformance index threshold of the neural network can balance the calculation speed and precision and it is benefit for the on-line identification of signals. The simulationexamples are analyzed in the background of random white noise superimposingdifferent signal to noise ratios. It showed that the method can effectively suppressvarious noise levels, and it effectively improves the anti-noise ability of Pronymethod.