Wide-area monitoring and recognition for power system disturbances using data mining and knowledge discovery (DMKD) theory.

摘要

This dissertation proposes a Wavelet Transform-based (WT) disturbance recognition approach based on Data Mining and Knowledge Discovery (DMKD) theory. The approach aims to process synchrophasor data from Wide-area Monitoring System (WAMS) to effectively recognize power system disturbances, such as generation loss, load change, and line trip, and discover knowledge about power system performance and WAMS based on recognition results.;Signature extraction and pattern recognition are two key steps in DMKD. In this dissertation, the approaches for signature extraction and pattern recognition are studied in detail. Compared with time-domain analysis, wavelet analysis, as a mature and promising tool of time-frequency-domain analysis, is proposed to extract disturbance signature from synchronous data. The criterion for selections of wavelet function and optimal decomposition scale is determined by maximum wavelet energy. Wavelet coefficients (WCs) at scale 5 obtained by order 2 Daubechies Wavelet are considered as disturbance signature. The WCs from measuring locations are combined to be a signature vector, an ordered number series, to reflect the penetration of disturbances in the entire power system. The signature vector is further compressed by Differential Box-Counting (DBC) method of Fractal Analysis (FA) to form a simplified signature for pattern recognition. Random Forest(TM) (RF) is chosen to be the classifier for pattern recognition. To achieve best recognition results, the parameter (number of trees) of RF(TM) and the training method for RF(TM) are investigated, respectively. Ten trees are generated to create an RF(TM) and ten-fold cross-validation technique is used to train RF(TM) and test the recognition accuracy. The recognition results achieve an overall 92% correction rate.;Based on extracted signatures and recognition results, some points of knowledge are discovered and discussed in this dissertation, including the correlation between WCs and load power variation, recognition for cascading line trip, signature characteristics in different generation and load levels and the influence of information redundancy of WAMS to recognition accuracy.