Using multiple data sources to detect manipulated electricity meter by an entropy-inspired metric

Abstract With the digitalization of electricity meters many previously solved security problems, such as electricity theft, are reintroduced as IT related challenges which require modern detection schemes based on data analysis, machine learning and forecasting. Here, we demonstrate a multidimensional anomaly detection approach for the early detection of tampered with electricity meters by comparing a set of multiple energy demand time series. Our method can complement and enhance existing monitoring systems which usually only analyze a single time series. We aim to detect electricity theft, which leads to noticeable outliers in our work. We present three data preprocessing methods to produce outliers in case of energy theft and highlight the requirements and fine-tuning mechanisms for the aggregation and comparison of multiple data sources. We show that our metric is robust against multiple manipulated data sources, which is a concrete improvement to alternative outlier preserving concepts to aggregate multiple data sources. With detection rates better than 90%, we demonstrate the effectiveness of using several data sources simultaneously, that, when used individually, provide little value in anomaly detection. Furthermore, we show that we can use different households as comparable data sources, without clustering the households according to their similarity first.