Extracting Most Discriminative Features on Transient Multivariate Time Series by Bi-Mode Hybrid Feature Selection Scheme for Transient Stability Prediction

Abstract

Real-time transient stability assessment (TSA) of power systems based on mining system dynamic response has been widely considered by scholars. In this regard, extracting the most discriminative transient features (MDTFs) to achieve high-performance transient stability prediction (TSP) should be regarded as a fundamental issue in the transient learning strategy. In fact, MDTFs extraction is raised to make a trade-off between paradoxically intertwined indices, namely the accuracy and processing time of TSP. To this end, we offer a bi-mode hybrid feature selection scheme called BMHFSS for extracting MDTFs in high dimensional transient multivariate time series (TMTS). First, we used the TMTS, which are effective features on TSA. Next, the trajectory-based filter-wrapper mode (TFWM) is applied on TMTS to surmount the curse of dimensionality in two phases. In the filter phase, statistical and intrinsic characteristics of the TMTS in the form of agglomerative hierarchical clustering (AHC) are measured, and relevant TMTS (RTMTS) is selected according to obtained weight. In the wrapper phase, the RTMTS is entered into the trihedral kernel-based approach, including both fuzzy imperialist competitive algorithm (FICA) and incremental wrapper subset selection (IWSS) to find the intersected most RTMTS (IMRTMTS). As a complementary step, the filter-wrapper scenario in point-based mode (PFWM) is conducted for selecting MDTFs per time series in IMRTMTS. Finally, the aggregated MDTFs (AMDTFs) are tested to verify their efficacy for TSP based on cross-validation. The results show that the proposed framework has prediction accuracy greater than 98 % and a processing time of 52.94 milliseconds for TSA.