Research On Features Extraction And Identification Methods For Power System Measured Overvoltages

Overvoltages is the main reason for the insulation damage of electrical equipment in power network, but also the determining factor in the design of the dielectric strength of electrical equipment and of transmission and distribution lines. The types of overvoltages occurred in the power system are various and their occurrence mechanism is different to each other. The amplitudes, frequencies and duration of overvoltages are very different, their damages to power system are also not the same. Therefore, automatic classification and identification of power system overvoltage types is the precondition to take overvoltages suppression measures. It is very important for real-time acquisition and automatic identification of the operation state of power grid to improve its self-healing capacity. Based on the measured overvoltages data of a 110 kV substation, the time-frequency domain characteristics and nonlinear dynamic behaviour of overvoltage signals, feature extraction and classification system construction for overvoltage signals are thoroughly researched, the main contents are as follows:For the problem of extracting features from switching overvoltages signals difficultly, based on the depth analysis on the time-domain and frequency-domain characteristics of four classes of switching overvoltage signals, combining with their generation mechanism and affecting factors, the characteristics which are distinctive, stable, and have explicit physical meaning are specified. The four types of switching overvoltages are caused by unloaded transformer energization, capacitor bank energization, unloaded line energization, unloaded line energization respectively. The RMS method, Fourier transform, wavelet transform, etc signal processing methods are integrated to extract their features. The research shows the features extracted for the 4 types of overvoltages are all distinctive and stable.In order to reduce the complexity of the classifiers and construct a smart identification system, a modularization tree-structured overvoltages identification system is built. A recognition module for identifying one type of overvoltages is built at each tree node. The tree node is divided into a leaf node and a branch node according to the output of the recognition module. Overvoltage signals will be pre-processed, extracted features and classified in the identification modules. The feature vector dimensions of the four types of overvoltages are 5, 8, 8, 3 respectively, all are much lower than that of their combination vector 16. Thereby the classifier complexity of each module is reduced. The total identification rate of the identification system for the four types and unknown types of measured overvoltage samples is 94.3%. Its effectiveness is verified.For the problem that nonlinear ferroresonance behaviour in power system is difficult to be identified by conventional linear mathematical methods, in the case of only a single measured voltage time series, in order to identifying accurately the ferroresonance types, the delay coordinates method is applied to reconstruct the system phase space. And then 3 nonlinear dynamics methods, e.g. phase plane, Poincaré section and correlation dimension, are used to express its dynamic characteristics. A comparative analysis of the dynamic characteristics of 3 typical time series with 4 data length is carried out, and their ferroresonance types are identified accurately. The required appropriate data length for applying the above 3 analysis methods and the applicability of the Cao’s method to obtain the minimum embedding dimension are discussed. The research shows that the measured voltage time series is affected by many factors, the identification result of ferroresonance mode occurred is more accurate and convictive through synthesizing the features characterized by the above 3 analysis methods; the data lengths required for different motion patterns and different analysis methods are all different; the D2 values estimated only in the minimum embedding dimension determined by Cao’s method may lead to erroneous conclusions.In order to solve the problem of feature interference that may exist when different types of overvoltages were mixed, a classification method of extracting features and combining multi-label based on time-domain segment division is proposed to identify mixed overvoltages. Firstly, based on wavelet energy spectrum, a algorithm to detect the starting and ending time of multiple transient fast decaying oscillations(TFDO) is proposed, the detect results obtained will be used to divide the overvoltage signals into multiple time-domain segment. The time-frequency domain characteristics of TFDOs are first extracted to identify the types of transients. The corresponding features are extracted from other time-domain segments to carry out their identification according to overvoltage types which may be induced by or coexisted with the identified transient types. The type of mixed overvoltages is obtained by combining multiple time-domain segment labels. By mining their general characters and relations of time-domain characteristics of multiple arc grounding transients contained in the analyzed data, the reasoning rules are built to identify a single arc grounding transient, increasing the identification rate of arc grounding transients. The research on multiple time-domain segments division, features extraction and identification about four categories of overvoltage signals mixed with arc grounding or lightning transients are carried out. The total identification rate of the constructed classification division tree for identifying mixed overvoltages is 94.6%, which shows that the proposed identification method is effective.To build a relatively complete overvoltages identification system, a method of constructing identification system for overvoltages based on subset hierarchy is proposed. Firstly, based on the difference between the three phase voltage RMS values of overvoltage signals and normal signals, the overvoltage category set is divided into two subsets: subset I and subset I’. The subset I’ is further subdivided into subset II and subset III by the windowed wavelet energy extracted. The overvoltage signals belong to subset II do not contain transients, while which belongs to subset III contain transients. Finally, the overvoltage signals classified to three subsets I, II and III are gradually identified applying the aforementioned modularization identification method. The overall identification rate of the constructed system for 8 categories of single overvoltages and 4 categories of mixed overvoltages is 97.1%. Its effectiveness is verified. It lays a foundation towards engineering application.