Application TadGAN to Detect Collective Anomaly in Power Usage Data

Abstract

In order to develop an efficient power generation plan, it is necessary to identify consumers’ power usage patterns. In general, power usage data takes the form of time series data and in order to analyze that data, it is necessary to first verify that there is no data contamination. To this end, the process of verifying that there are no anomalies in the data is essential. In particular, for power data, anomalies are often recorded across multiple time units rather than just one point. In this work, we applied the TadGAN algorithm to detect these collective anomalies. Using the power usage data recorded in the actual building, the anomalies were injected randomly with various conditions. Afterwards, we detected anomalies by using TadGAN and showed that we were better at detecting collective anomaly than point anomaly.