Machine learning based real-time monitoring of long-term voltage stability using voltage stability indices
Loading...
Date
2020-12-24
Authors
Dharmapala, Kalana
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Voltage instability manifests as a progressive voltage drop leading to a voltage collapse. It has been associated with many recent blackouts occurred at various parts of the world. Ability to predict the proximity to voltage instability in real-time using measurements can improve the security of power systems.
Among the various techniques that are available to assess the proximity to voltage instability of a power system, Voltage Stability Indices (VSIs) are the widely considered as suitable for generating real-time quantitative parameters that indicate the voltage security level. The majority of the VSIs are theoretically coherent at the proximity of the point of system voltage collapse where they reach some critical value. However, their ability to accurately indicate the voltage stability margin varies with power system operating conditions and under different contingencies: It becomes difficult for the system operators to differentiate voltage instabilities from normal operating conditions using simple rules such as thresholds.
Loadability Margin (LM) is a direct representation of proximity to long-term voltage stability and it is also an easily understandable indicator of voltage stability. However, it is difficult to determine the LM of a large power system through direct computations in real-time. As a solution, a new machine learning based approach is suggested to predict the LM. The proposed technique uses different VSIs computed from synchrophasor measurements as inputs to several Machine Learning Models (MLMs) which forms an ensemble that collectively predict the LM. The VSIs used for each MLM is carefully chosen to include those based on different principles. A procedure that include automated training data generation, correlation based input feature selection, and performance based choice of machine learning algorithm is proposed to develop the proposed system. The studies conducted on the IEEE-14 and -118 bus systems showed that the Random Forest Regression gives the best performance in terms of the accuracy and robustness.
The effects of practical aspects such as synchrophasor measurement errors on the LM prediction accuracy were analyzed. Finally, a real-time version of the proposed LM prediction system was implemented on PhasorSmart® synchrophasor application platform and was tested using the IEEE 14-bus system and phasor measurement units simulated on RTDS® real-time simulator.
Description
Keywords
Long term voltage stability, Voltage stability indices, Continuation power flow, Synchrophasors, Machine learning based voltage stability assessment