Predicting the Power Level of a Nuclear Reactor using a Time Series-based Approach

摘要

Detecting the power level of a nuclear reactor and in turn verifying that it is operating at the declared power level is a problem of interest for the nuclear nonproliferation community. We use data collected from multiple sensor modalities (seismic, acoustic, effluent, electromagnetic, and thermal) positioned near a collocated research nuclear reactor and reprocessing facility at Oak Ridge National Laboratory for the Multi-Informatics for Nuclear Operations Scenarios (MINOS) venture. Naive Bayes, a classification method which is robust against missing data, has demonstrated greater predictability than random forest for the power level of a reactor. However, this classification method assumes samples are independent, which has been found prone to rapid variations and difficulties in predicting short-term power holds. Realistically, there is temporal dependency in the reactor power, i.e., reactor power cannot change by a large amount in a short period and the reactor is more likely to stay at the previous power level. In this paper, we explore and compare two extensions of the Naive Bayes classifier (naive Bayes sequential and hidden Markov model) that take into account this temporal dependency as well as an ensemble model that combined predictions from all three models.