Data-Driven Methods for Managing Anomalies in Energy Time Series

With the progressing implementation of the smart grid, more and more smart meters record power or energy consumption and generation as time series. The increasing availability of these recorded energy time series enables the goal of the automated operation of smart grid applications such as load analysis, load forecasting, and load management. However, to perform well, these applications usually require clean data that describes the typical behavior of the underlying system well. Unfortunately, recorded energy time series are usually not clean but contain anomalies, i.e., patterns that deviate from what is considered normal. Since anomalies thus potentially contain data points or patterns that represent false or misleading information, they can be problematic for any analysis of this data performed by smart grid applications.

Therefore, the present thesis proposes data-driven methods for managing anomalies in energy time series. It introduces an anomaly management whose characteristics correspond to steps in a sequential pipeline, namely anomaly detection, anomaly compensation, and a subsequent application. Using forecasting as an exemplary subsequent application and real-world data with inserted synthetic and labeled anomalies, this thesis answers four research questions along that pipeline for managing anomalies in energy time series.
... mehr
Based on the answers to these four research questions, the anomaly management presented in this thesis exhibits four characteristics. First, the presented anomaly management is guided by well-defined anomalies derived from real-world energy time series. These anomalies serve as a basis for generating synthetic anomalies in energy time series to promote the development of powerful anomaly detection methods. Second, the presented anomaly management applies an anomaly detection approach to energy time series that is capable of providing a high anomaly detection performance. Third, the presented anomaly management also compensates detected anomalies in energy time series realistically by considering the characteristics of the respective data. Fourth, the proposed anomaly management applies and evaluates general anomaly management strategies in view of the subsequent forecasting that uses this data. The comparison shows that managing anomalies well is essential, as the compensation strategy, which detects and compensates anomalies in the input data before applying a forecasting method, is the most beneficial strategy when the input data contains anomalies.