As a power system evolves in size and complexity due to energy transition, the focus on the stability of the power system becomes more imperative than ever. Power system stability is tested when the system is subjected to large disturbances. Consequently, a conventional protectio
...
As a power system evolves in size and complexity due to energy transition, the focus on the stability of the power system becomes more imperative than ever. Power system stability is tested when the system is subjected to large disturbances. Consequently, a conventional protection scheme is always deployed to take quick and automatic actions based on localized measurements, to restore the power system to a steady operating state. However, in many cases, further remedial actions by the system operator are necessary to avoid cascading failures and eventual instability problems. The introduction of phasor measurement units in the form of WAMS has led to improved awareness and
hence, better, capability to take remedial actions. In particular, the vast amount of data provided by the PMUs have led to the development of data-driven machine learning applications to aid the system operator in maintaining stability. One such application is power system event classification. In general, event classification is achieved using supervised and unsupervised methods. A review of pioneering literature reveals that offline models trained using both these methods can achieve a high order of classification accuracy. However, in a realistic, online setting with streaming PMU data, there is a possibility
of new, unknown events appearing in the data stream. In such a case, Offline
models which are trained on a limited number of known classes are unable to adapt to new class data. Therefore, there is a necessity for event learning applications that are capable of incrementally identifying new classes that appear in the PMU data stream. Hence, this thesis proposes the event clustering algorithm which performs in two stages – label correction stage and new class labeling stage. In the first stage, the algorithm corrects the class labels predicted by a supervised model, and flags and collects new incoming event instances through a shape-based similarity measure, i.e. DTW measure. In the subsequent second stage, it incrementally clusters the collected new event instances using Time-Series k-means clustering, where the hyperparameter of the number of clusters is automatically determined using a silhouette score clustering metric. The results of this algorithm show successful detection of new power system events with high classification accuracy. Also, final improvements are suggested to overcome certain drawbacks of the algorithm listed in the thesis.
