Power Oscillation Monitoring using Statistical Learning Methods

This paper describes development and testing of a method for estimation of electromechanic modes and corresponding mode shapes from frequency or voltage angle measurements in Wide Area Monitioring Systems. The method uses Complex Principal Component Analysis to perform a decomposition of the dynamics captured in the measurements, reducing any detected oscillations into sets of parameters. The numerous sets of parameters generated after a time period can be interpreted as points or observations in high-dimensional space, on which a clustering algorithm can be applied to pinpoint areas where high densities of points are accumulated. Results show that highly accurate estimates of oscillatory modes in the power system and their mode shapes can be derived by averaging the point observations belonging to each cluster. Important development described in this paper includes the use of voltage angle as input and testing on a medium size power system simulation model with synchronized measurements from 44 nodes. Furthermore, introduction of the DBSCAN clustering algorithm shows very promising results when applying the method to recorded phasor measurements from the Nordic power system.