Most of us only think about voltage when we replace a battery in an appliance such as a child’s toy, electric toothbrush, or remote. On a larger scale, voltage levels vary for different sections of power systems and has to be kept within a set range for the power system to work properly and safely. Too high of voltage can cause equipment to be damaged or fried while low voltage can cause equipment to operate poorly or stall. The impact of power system instability (e.g., voltage instability, frequency stability) can be seen in the increase of outages in recent years. In the past decade alone, power outages have nearly tripled to affect 37 million people annually.

Professors Nanpeng Yu of electrical and computer engineering and Eamonn Keogh of computer science and engineering received a $1 million award from the Department of Energy to identify power system stability problems and other anomalies in real-time. By developing advanced machine learning and data mining algorithms, the team will produce algorithms and software packages critical to enhancing the wide-area situational awareness, visualization, protection, and control of large-scale power systems.

The team’s objective is three-fold. First the team will identify anomalous events using innovative time series data mining technology Matrix Profile and corresponding algorithms developed by the project team. The data gathered will help detect anomalous events hidden in high-dimensional synchrophasor data and discover signatures (including precursors signatures) for various events/dynamic patterns in the national power grid.

The team will then use data labels such as asset failure, anomaly events, and instability category to create a catalog of event signatures. The resulting effort will train deep neural networks, which are capable of predicting asset health and learning precursors to different instability problems in the three interconnections.

The final objective will predict asset health and learn precursors to instability phenomenon. The team will implement the Matrix Profile technology and the deep recurrent neural network on GPU platforms to finish mining the large scale synchrophasor data, which spans over two years.

In terms of wide-area monitoring, unsupervised and supervised machine learning, algorithms can easily identify power system oscillations in an area and identify static and dynamic voltage stability problems in real-time. Regarding wide area protection and control, the proposed algorithms are capable of predicting and detecting devastating events such as transient instability. The quick recognition of potential instability and unintentional islanding, and automatic initiation of emergency control are critical for preventing widespread outages.

The advanced data mining and machine learning algorithms will greatly benefit independent system operator and electrical utilities companies in the United States. These beneficiaries can either directly download the open-source version of the algorithm or access them through commercial software packages, which will integrate the advanced data mining and machine learning algorithms.