Synchrophasor Data Mining for Situational Awareness in Power Systems

Recently, there has been an increase in the deployment of Phasor Measurement Units (PMUs) which has enabled real time, wide area monitoring of power systems. PMUs can synchronously measure operating parameters across the grid at typically 30 samples per second, compared to 1 sample per 2-5 seconds of a conventional Supervisory Control And Data Acquisition (SCADA) system. Such an explosion of data in power systems has provided an opportunity to make electrical grids more reliable. Additionally, it has brought a challenge to extract information from the massive amount of data. In this research, several data mining algorithms are used to extract information from synchrophasor data for improving situational awareness of power systems. The extracted information can be used for event detection, for reducing the dimension of data without losing information, and also to use it as heuristic to process future measurements. The methods proposed in this research work can be broadly classified into two parts: a) stream mining and b) dimension reduction. Stream mining algorithms provide solution utilizing state-of-the-art data stream mining algorithms such as Hoeffding Trees (HT). HT algorithm builds a decision tree by scanning the incoming data stream only once. The tree itself holds sufficient statistics in its leaves to grow the tree and also to make classification decisions of incoming data. Instead of using a large number of samples, which leads to a tree too large to accommodate in memory, the number of samples that are needed to split at each node is determined using Hoeffding bound (HB). HB keeps the size of the decision tree within bounds while also maintaining accuracies statistically competitive to traditional decision trees. Dimension reduction algorithms reduce dimension of the synchrophasor data by extracting maximum information from a huge data set without losing information. In this dissertation, both online and offline dimension reduction algorithms have been studied. The online dimension reduction uses an unsupervised method using principal components of the time series data. The offline method optimizes unique mutual information between the state of the power system and synchrophasor measurements. It optimizes the criteria by reducing redundant information while maximizing relevant information