PaperlayerView exact revision on arXiv

Source-linked AI summary

Application of EOS-ELM with binary Jaya-based feature selection to real-time transient stability assessment using PMU data

Yang Li, Zhen Yang

arXiv:1809.02859v1eess.SP

TL;DR

Existing PRTSA methods can require excessive training time and parameter tuning, limiting real-time implementation and online model updating. This paper combines EOS-ELM with BinJaya feature selection using PMU data. On the IEEE 39-bus and provincial systems, it reports superior speed and accuracy to other sequential learning algorithms and reduces input dimension to about one-third without sacrificing classification performance.

  • Problem

    Many PRTSA methods require excessive training time and complex parameter tuning, limiting real-time implementation and online model updating.

    Source: §I · pp. 1–2

  • Method

    The paper proposes PMU-based PRTSA using EOS-ELM with BinJaya feature selection.

    Source: §I · p. 2, §VI · p. 8

  • Results

    The proposal has superior computation speed and prediction accuracy to SGBP, GAP-RBF, and MRAN, while BinJaya reduces input dimension to about one-third without sacrificing classification performance.

    Source: §VI · p. 8

  • Takeaways & Limitations

    Ensemble learning improves the stability and generalization ability of original OS-ELM.

    Source: §VI · p. 8

Abstract

from arXiv · show

Recent studies show that pattern-recognition-based transient stability assessment (PRTSA) is a promising approach for predicting the transient stability status of power systems. However, many of the current well-known PRTSA methods suffer from excessive training time and complex tuning of parameters, resulting in inefficiency for real-time implementation and lacking the online model updating ability. In this paper, a novel PRTSA approach based on an ensemble of OS-extreme learning machine (EOSELM) with binary Jaya (BinJaya)-based feature selection is proposed with the use of phasor measurement units (PMUs) data. After briefly describing the principles of OS-ELM, an EOS-ELM-based PRTSA model is built to predict the post-fault transient stability status of power systems in real time by integrating OS-ELM and an online boosting algorithm, respectively, as a weak classifier and an ensemble learning algorithm. Furthermore, a BinJaya-based feature selection approach is put forward for selecting an optimal feature subset from the entire feature space constituted by a group of system-level classification features extracted from PMU data. The application results on the IEEE 39-bus system and a real provincial system show that the proposal has superior computation speed and prediction accuracy than other state-of-the-art sequential learning algorithms. In addition, without sacrificing the classification performance, the dimension of the input space has been reduced to about one-third of its initial value.

I. INTRODUCTION

The introduction motivates fast, accurate post-fault transient stability assessment and identifies limits in conventional pattern-recognition approaches. It then introduces OS-ELM’s sequential learning principles as background for the proposed approach.

I. INTRODUCTION

  • TSA must rapidly determine post-fault stability; pattern recognition maps system-operation features to stability status, while PMUs provide synchronized real-time measurements.Transient stability can require corrective action within <1 s.Source: §I · p. 1
  • Conventional PRTSA can require excessive training and parameter tuning, while offline samples may not represent changing operating modes and special cases.When a model performs poorly on an online sample, it may need termination and offline retraining.Source: §I · pp. 1–2
  • OS-ELM learns one sample or chunk at a time and discards data after training, while ensemble learning combines multiple learners’ results.Feature selection is also presented as a way to support faster classification with comparable or better generalization.Source: §I · p. 2

II. PRINCIPLES OF OS-ELM

  • OS-ELM initializes from an initial training chunk with randomly generated hidden-node parameters, then processes new observation chunks and determines updated output weights.The paper outlines initialization and sequential-learning phases.Source: §II · p. 2

A. EOS-ELM

EOS-ELM uses OS-ELM as a weak classifier and online boosting as the ensemble mechanism. Selectors choose weak classifiers and combine their outputs into a strong classifier.

A. EOS-ELM

  • OS-ELM serves as the weak classifier, while an online boosting algorithm provides the ensemble-learning mechanism for EOS-ELM.The paper introduces this combination to address volatile OS-ELM output errors caused by randomly assigned hidden-node parameters.Source: A. EOS-ELM · p. 3
  • Each selector chooses the weak classifier with the lowest estimated error, and successive selectors receive updated sample importance weights.The chosen classifier’s estimated error determines its selection.Source: A. EOS-ELM · p. 3
  • A strong classifier combines selected weak classifiers linearly, and the selectors can correspond to features to select a subset from the global pool.This links the boosting selectors to feature-subset selection.Source: A. EOS-ELM · p. 3

B. EOS-ELM-based TSA Model

The proposed TSA model retains offline training and online assessment while adding online learning to update the model from samples reflecting current operating modes. Real-time prediction is triggered by a relay-protection tripping signal after a large disturbance.

B. EOS-ELM-based TSA Model

  • A knowledge base is built from time-domain simulations covering typical contingencies, with operating points represented by stability indices and input features.The knowledge base’s completeness and representativeness are described as important to model generalization.Source: B. EOS-ELM-based TSA Model · p. 3
  • The approach has offline-learning, online-learning, and real-time-prediction modes; online learning updates model structure and parameters when a new special case occurs.Online samples are obtained from simulation data reflecting current operating modes through the online DSA module of EMS.Source: B. EOS-ELM-based TSA Model · p. 4
  • After a large disturbance, a relay-protection tripping signal triggers real-time prediction using the trained TSA model’s mapping relationships.This contrasts with the conventional framework, where a poorly performing online model must be terminated and retrained offline.Source: Figure 1 · p. 3, B. EOS-ELM-based TSA Model · p. 4

A. Construction of the Original Features

The original feature set combines static and post-fault dynamic information from PMU data. It comprises system-level classification features, including measurements at specified cycles after a fault.

A. Construction of the Original Features

  • The feature design uses post-fault dynamic information alongside static information to make use of PMU data.Earlier work mainly used pre-fault static features because SCADA measurements could not provide wide-area post-fault dynamic information.Source: A. Construction of the Original Features · p. 4
  • The original feature set consists of system-level classification features, with tcl+3c, tcl+6c, and tcl+9c denoting the 3rd, 6th, and 9th cycle after the fault.Table I is identified as listing the original input features.Source: A. Construction of the Original Features · p. 4, Table I · p. 4

B. BinJaya-based Feature Selection

The section presents BinJaya with KFRS for selecting features and describes how Jaya is adapted to generate binary feature subsets.

B. BinJaya-based Feature Selection

1) Class Separability Criterion

  • KFRS generalized classification function gc(D) serves as a class-separability criterion, combining generalized dependency and classification-certainty functions over feature subset B.Source: B. BinJaya-based Feature Selection · p. 4

2) Jaya algorithm

3) Angle Modulation

V. CASE STUDY

The case study evaluates the proposal on the IEEE 39-bus system and a real provincial system in China.

V. CASE STUDY

  • Simulations tested the IEEE 39-bus system and a real provincial system in China using MATLAB on a PC with two 2.4 GHz Intel Core dual-core CPUs and 6 GB RAM.Source: §V · p. 5

A. Case 1—IEEE 39-bus System

The IEEE 39-bus case evaluates EOS-ELM using simulated fault scenarios and a BinJaya-selected feature subset. EOS-ELM improves on the original OS-ELM and compares favorably with other sequential learners, with results assessed across accuracy, stability, and training time.

A. Case 1—IEEE 39-bus System

  • The simulated knowledge base contains 3,300 samples, with 2,200 used for training and the remainder for testing.Scenarios include three-phase-to-ground faults at 60 locations, 11 load levels from 80% to 130% of base load, and five randomly assigned generator-output settings per level.Source: Figure 4 · p. 5
  • The OS-ELM validation accuracy peaked at 98.50% with 65 hidden nodes, so 65 was selected for this testing case.The number of hidden nodes was selected by cross-validation to maximize validation accuracy.Source: Figure 7 · p. 6, Figure 6 · p. 6
  • With 10 weak classifiers, EOS-ELM achieved the best predictive performance among the tested ensemble sizes, combining the lowest SD with better testing accuracy than the alternatives.The ensemble size was selected using results from 50 trials; the weak classifiers use the same hidden-node count as the original OS-ELM.Source: A. Case 1—IEEE 39-bus System · p. 6, Table II · p. 6
  • EOS-ELM increased testing accuracy by 1.14% and decreased testing SD by 0.0154 versus original OS-ELM, while training time was the exception to its otherwise better indicators.The comparison uses the BinJaya-selected subset OFS1 = {Tz4, Tz9, Tz19, Tz25, Tz26, Tz31, Tz32}.Source: A. Case 1—IEEE 39-bus System · pp. 6–7, Table III · p. 7
  • EOS-ELM had far better accuracy and stability than GAP-RBF and MRAN with much lower training time; versus SGBP, it had higher testing accuracy and lower SD but slightly longer training time.The comparisons use one-by-one learning; Table IV reports averages over 50 runs.Source: Table IV · p. 7, A. Case 1—IEEE 39-bus System · p. 7

B. Case 2—Real Power System of Liaoning Province

The Liaoning provincial-system case tests the approach on a large modeled grid using simulated fault scenarios. OFS2 outperforms OFS1 in prediction performance, and the authors relate the larger system’s increased stability-pattern complexity to the need for more input features.

B. Case 2—Real Power System of Liaoning Province

  • The modeled Liaoning system has 91 generators and 750 major buses, with five connected channels to the external network.The system covers 148,000 square kilometers and has about 39,657.2 MW of installed capacity.Source: B. Case 2—Real Power System of… · p. 7
  • The test set was drawn from 2,000 simulated samples after 1,320 were assigned to training, across load levels from 80% to 130% of basic load.Scenarios used three-phase-to-ground faults cleared after five to ten cycles at several line locations; OFS2 contains 12 features, and the selected EOS-ELM has 15 OS-ELM networks.Source: B. Case 2—Real Power System of… · p. 7
  • OFS2 has better prediction performance than OFS1 in the Liaoning case; Table V reports the comparison using accuracies averaged over 50 runs.The authors attribute the advantage to the greater stability-pattern complexity of larger systems, which calls for more input features.Source: B. Case 2—Real Power System of… · pp. 7–8, Table V · p. 8

VI. CONCLUSIONS

The conclusions report that EOS-ELM with BinJaya-based feature selection compares favorably with other sequential learning algorithms and reduces input dimensionality. They also identify ensemble learning’s reported benefits and suggest possible future applications.

VI. CONCLUSIONS

  • The proposal has superior computation speed and prediction accuracy to SGBP, GAP-RBF, and MRAN.Source: §VI · p. 8
  • BinJaya reduces input-space dimension to about one-third of its initial scale without sacrificing classification performance.Source: §VI · p. 8
  • Ensemble learning techniques are reported to improve the stability and generalization ability of the original OS-ELM.Source: §VI · p. 8
  • Future work may use the proposal to trigger wide-area protection and control, and may apply BinJaya feature selection to similar engineering classification problems.Source: §VI · p. 8
Loading 1809.02859v1…