This project aims to evaluate the feasibility of developing a virtual sensor network (VSN) for inferring transmission line conditions as specific points based on data gathered from other points in the system. The North American electric grid interconnections arguably constitute the largest machines on the planet, and millions of people depend on them for safety, security, and comfort. The reliability of these systems is crucial to health care, business, education, and every other aspect of economic activity, but sensor penetration in the network is very low.
The Western Electricity Coordinating Council (WECC) is currently leading the Western Interconnection Syncrophasor Program to install more han 300 new or upgraded Phasor Measurement Units (PMUs) to collect real-time data in an effort to increase grid reliability at a cost of about 107.8 million. With the successful development of a virtual sensor network, those 300 sensors could be redistributed over a wider area, ncreasing data collection at a fraction of the cost of purchasing, installing, and maintaining new sensors. To develop a VSN, data is acquired from established sensors and fed into a neural network that establishes which measurements can be inferred from remote data. Redundant sensors can then be redistributed to expand the sensing network. A cursory review of peer-reviewed literature reveals that this technique has been applied with some success to wind-speed prediction on wind farms; however, the application to electric transmission systems does not seem to have yet been made.
To evaluate the feasibility of a VSN for transmission systems, software models of transmission systems are created or imported to appropriate software, such as PowerSim in Matlab or Alternative Transients Program (ATP), to develop a collection of data on system states and responses to transient events. This data is then provided to a neural network program developed using the Neural Networking Toolbox in Matlab to establish which data points are required and which can be predicted. The neural network is then evaluated by changing conditions in the system model, feeding selected data into the neural network, and comparing the outputs to the states described in the model.
