Power Quality Abnormal Events Online Detection And Diagnosis Technology

With the gradual and energetic development of the Chinese power system,the structure of the power grid becomes increasingly complex,and the types of connected circuit elements are different.Among them,the fluctuation and abnormality caused by nonlinear and impulsive structural monomer are very significant.Harmonic pollution is an important cause of abnormal power quality.Harmonic,inter-harmonic and sub-harmonic have an impact on various indicators of power quality at all times.The resulting abnormalities include three-phase unbalance,voltage sag,voltage flicker,frequency deviation,etc.According to the duration and abnormal shape,these anomalies can be divided into steady-state anomalies and transient anomalies.It is of great significance to accurately detect and analyze the causes of abnormal events.However,because power quality data often has the characteristics of high dimension and high noise,it is also a challenge to carry out data processing and feature extraction reasonably without affecting the recognition accuracy.In addition,the power index is also cyclical,and abnormal conditions need to be detected and analyzed in the time sequence dimension.Based on this background,this paper focuses on the detection of abnormal power quality events and the identification and positioning of abnormal causes,and has made the following work:1.For anomaly detection of power quality data,this paper proposes a power quality anomaly detection algorithm based on Gaussian tail probability thresholding.Based on traditional statistical control chart algorithms,this algorithm calculates the anomaly likelihood value and uses Gaussian tail probability density to measure the anomaly score distribution,replacing the traditional direct thresholding method,effectively improving the noise resistance of the model,with an anomaly detection accuracy of around 90%.2.For the classification and recognition of power abnormal events,traditional signal algorithms rely heavily on manual feature extraction with business experts.Although deep learning methods do not require manual feature extraction,due to the characteristics of high power quality noise,it is prone to over fitting.In view of the current situation,this paper proposes a de-operated power quality disturbance identification and diagnosis model based on random forests,which improves the combination of the Relief F feature extraction algorithm.It can accurately identify the causes of power quality disturbances without the need for business experts.Compared with traditional methods,it has strong noise resistance and accuracy up to 89%.3.In order to solve the difficult problem of locating disturbances in distributed power grids with complex structures,this paper uses an Elastic-Net network to determine the impact of different measurement points on abnormal events,and combines the Pearson correlation coefficient to locate interference sources to locate and analyze the occurrence of anomalies.