LSTM-based quick event detection in power systems

In this paper, a data-driven online approach is established to detect events in power systems in real time. The approach does not require prior knowledge of the power system model or its parameters. Instead, it utilizes a long short-term memory (LSTM) model to capture the state evolution of the power system. Due to the expressiveness of the LSTM model, it is able to track the system states with small prediction error when it operates under normal conditions. However, when the system is perturbed by certain events that cannot be predicted by the model, the prediction error will increase dramatically. Thus, by tracking the prediction error of the trained LSTM model, the data-driven online approach is able to detect events in a timely fashion. The event detection problem is then cast into the quick change detection framework, where a Cumulative Sum (CUSUM) based approach is proposed. To overcome the difficulty that the statistics of the prediction error when events happen is generally unknown beforehand, a generalized likelihood ratio test (GLRT) is incorporated into the CUSUM procedure. A Rao-test is then adopted to reduce the computationally complexity of GLRT. Finally, the LSTM based event detection approach is validated with real-world PMU measurements.

Published in: 2020 IEEE Power & Energy Society General Meeting (PESGM)

Files

Metadata

Work Title LSTM-based quick event detection in power systems
Access
Open Access
Creators
  1. Boyu Wang
  2. Yan Li
  3. Jing Yang
Keyword
  1. Event detection
  2. LSTM
  3. CUSUM
License In Copyright (Rights Reserved)
Work Type Conference Proceeding
Publication Date December 16, 2020
Publisher Identifier (DOI)
  1. https://doi.org/10.1109/PESGM41954.2020.9281569
Deposited March 18, 2024

Versions

Analytics

Collections

This resource is currently not in any collection.

Work History

Version 1
published

  • Created
  • Added LSTM-based_Quick_Event_Detection_in_Power_Systems.pdf
  • Added Creator Boyu Wang
  • Added Creator Yan Li
  • Added Creator Jing Yang
  • Published
  • Updated Keyword, Description, Publication Date Show Changes
    Keyword
    • Event detection, LSTM, CUSUM
    Description
    • <p>In this paper, a data-driven online approach is established to detect events in power systems in real time. The approach does not require prior knowledge of the power system model or its parameters. Instead, it utilizes a long short-term memory (LSTM) model to capture the state evolution of the power system. Due to the expressiveness of the LSTM model, it is able to track the system states with small prediction error when it operates under normal conditions. However, when the system is perturbed by certain events that cannot be predicted by the model, the prediction error will increase dramatically. Thus, by tracking the prediction error of the trained LSTM model, the data-driven online approach is able to detect events in a timely fashion. The event detection problem is then cast into the quick change detection framework, where a Cumulative Sum (CUSUM) based approach is proposed. To overcome the difficulty that the statistics of the prediction error when events happen is generally unknown beforehand, a generalized likelihood ratio test (GLRT) is incorporated into the CUSUM procedure. A Rao-test is then adopted to reduce the computationally complexity of GLRT. Finally, the LSTM based event detection approach is validated with real-world PMU measurements. </p>
    • <p>In this paper, a data-driven online approach is established to detect events in power systems in real time. The approach does not require prior knowledge of the power system model or its parameters. Instead, it utilizes a long short-term memory (LSTM) model to capture the state evolution of the power system. Due to the expressiveness of the LSTM model, it is able to track the system states with small prediction error when it operates under normal conditions. However, when the system is perturbed by certain events that cannot be predicted by the model, the prediction error will increase dramatically. Thus, by tracking the prediction error of the trained LSTM model, the data-driven online approach is able to detect events in a timely fashion. The event detection problem is then cast into the quick change detection framework, where a Cumulative Sum (CUSUM) based approach is proposed. To overcome the difficulty that the statistics of the prediction error when events happen is generally unknown beforehand, a generalized likelihood ratio test (GLRT) is incorporated into the CUSUM procedure. A Rao-test is then adopted to reduce the computationally complexity of GLRT. Finally, the LSTM based event detection approach is validated with real-world PMU measurements. </p>
    • Published in: 2020 IEEE Power & Energy Society General Meeting (PESGM)
    Publication Date
    • 2020-08-01
    • 2020-12-16
  • Updated