Research On Analysis Method Of Power Quality Based On Long Short-term Memory Networks

With the development of power electronic technology and microgrid technology,various nonlinear power electronic devices and surge loads have appeared in the power system,causing a series of power quality disturbance(PQD)phenomena.The deterioration of power quality has had a negative impact on the power system.Therefore,classifying and detecting PQD signal is a prerequisite for power system fault analysis.The traditional way of PQD detection has problems such as poor noise immunity,poor positioning.For these reasons,this paper studies the method of classifying and detecting PQD signal using long short-term memory network(LSTM).The main research work is as follows:The research progress of PQD detection and classification methods at home and abroad are analyzed,signal processing methods based on time-frequency domain and machine learning classification algorithms are discussed,and the commonly used feature extraction and classification methods are summarized.Aiming at the problem that PQD positioning is greatly affected by noise and complicated calculations,a PQD detection method based on the time-domain residuals of the LSTM network is proposed.The regression model of LSTM is used to estimate the voltage signal,and the difference between actual value and estimated value is used to obtain the residual.After the residual signal is filtered by the threshold,the mutation position can be obtained,and the starting and ending instants can be located.After the estimated value and the actual value is matched with the zero-crossing point,the phase change can be detected according to the displacement degree of the waveform displacement.Calculating the amplitude of the residual signal can detect changes in the amplitude of the voltage signal.The results of experiments show that the method has good noise immunity,and can locate the instants of sag,swell,and oscillatory transient accurately.The results also show that the method has high detection accuracy for changes in voltage amplitude and phase.The feature extraction step in PQD classification is affected by human factors and noises,which leads to poor versatility of features.Therefore,a PQD classification method based on long short-term memory fully convolutional network(LSTM-FCN)is proposed.In this method,the disturbed signal will be input into a classification model that composed of LSTM network and full convolutional network(FCN)in parallel,and the features extracted by the two types of network will be merged.The disturbance type will be automatically identified finally.The test results of simulated data show that the method has a high recognition rate for PQD signal,and has strong adaptability to noisy signal.The results of modeling and testing the measurements prove the method can be applied to power system fault analysis.In order to further extract the hidden information of the feature,without changing the structure of the LSTM-FCN model,an attention mechanism is added to the LSTM layer to filter the features,and then the filtered features are merged with the features extracted by FCN.The model will recognize the fusion features automatically.Experiment shows that the model can learn the characteristics more effectively after adding the attention mechanism.Attention mechanism enables the model to classify events more accurately and improves stability.