Abstract
Emergency detection is of significant value in preventing damages to power systems and even saving lives by alerting anomalous behaviors of electronic devices. However, existing works only identify manually encoded events or patterns on power supply and thus consume huge amount of manpower and cannot deal with undefined emergency. In this paper, we propose a novel method 3D-LRT (Three Dimensional Likelihood Ratio Test) to detect emergency in large scale power system. To the best of our knowledge, this is the first work that leverages spatiotemporal behavior characteristics to identify anomalous patterns in power system. For scalability of large scale power systems, we further optimize our 3D-LRT using pruning and parallelization methods to save time overhead. We conduct experiments on real-world synthetic data sets. The results demonstrate that our 3D-LRT method is both effective and efficient.