High Frequency Dynamic Voltage Forecasting Under Reduced Wireless Sensor Network Observability

Missing information in Wireless Sensor Network nodes can lead to inaccurate forecasts when state estimation is done on any particular system. This poses a serious threat to systems such as power distribution grids as a small miscalculation in the system can lead to disastrous effects. This paper proposes a relatively simple, yet novel, approach to forecasting missing information in a Wireless Sensor Network that collects high-frequency data. Specifically, the voltage measurements of the microPhasor Measurement Units from the Lawrence Berkeley National Lab with sampling frequency of 250Hz is analyzed. We model the problem as a matrix completion problem - similar to how collaborative filtering problems work. The average forecasting root mean square error is computed, estimated through the Frobenius Norm of the difference between the completed and the test matrices, at 0.03% while the standard deviation is around 0.15. We prove that there is a logarithmic increase of error when more data is missing - attributed to when the sensors are down more often, but the increase of error is negligible since it is still around 0.1%. The accurate measurements can be attributed to the high-frequency sampling rate used when collecting information - justifying the need for more high-frequency information collection through wireless sensor networks.