Abnormal Detection of Power Consumption Based on A Stacking Ensemble Model

Non-technical loss (NTL) caused by abnormal power consumption behavior has a negative impact on power quality and electrical utilities' economic benefits. In view of the accumulation of consumption data, many data-driven methods especially machine learning (ML) are applied to identify abnormal consumption behaviors, so this paper proposes a stacking ensemble model for abnormal consumption detection combining five different ML models. Incomplete consumption data is processed by two-step interpolation and consumption feature sets are established based on statistical analysis, time series characteristics and periodic similarity. Experiment results reveal that the proposed stacking model outperforms individual models and other combinations under four evaluation criteria. This approach has shown great performance in abnormal power consumption identification and thus can support to distinguish NTL for power monitoring and economic benefit management.