Predicting Power Failures with Reactive Point Processes

Abstract

We present a new statistical model for predicting discrete events continuously in time, called Reactive Point Processes (RPP’s). RPP’s are a natural fit for many domains where time-series data are available, and their development was motivated by the important problem of predicting serious events (fires, explosions, power failures) in the underground electrical grid of New York City (NYC). RPP’s capture several important properties of this domain: •There is an instantaneous rise in vulnerability to future serious events immediately following an occurrence of a past serious event, and this rise in vulnerability gradually fades back to the baseline level. This is a type of self-exciting property. •There is an instantaneous decrease in vulnerability due to an inspection or repair. The effect of this inspection fades gradually over time. This is a self-regulating property. •The cumulative effect of events or inspections can saturate, ensuring that vulnerability levels never stray too far beyond their baseline level. This captures diminishing returns of many events or inspections in a row.