Comparison of Enhanced Isolation Forest and Enhanced Local Outlier Factor in Anomalous Power Consumption Labelling

Abstract

Anomaly detection in power consumption is one of the major challenges faced by the modern world in response to the excessive electric consumption in developing countries. As a result, researchers were motivated to conduct extensive studies in this area to develop algorithms that classify the abnormal data instances from smart meter readings. In this paper, we examine and compare the effectiveness of two anomaly labelling algorithms, namely: the Enhanced Isolation Forest (E-IF) and the Enhanced Local Outlier Factor (E-LOF), in detecting the abnormal power consumption in building. The E-IF and the E-LOF are proposed based on the Isolation Forest (IF) and the Local Outlier Factor (LOF) algorithms with an additional step of applying a threshold to distinguish the high and low electricity consumptions anomalies. Experiments were performed to 10 smart meters readings and the capabilities of E-IF and E-LOF in detecting the injected anomalies were investigated. The results showed that the E-IF outperformed E-LOF, with E-IF managed to detect 100% of the injected anomalies at contamination levels of 0.30 and 0.35. The E-LOF, on the other hand, could detect an average of 68% of the injected anomalies for contamination level of 0.30 and 78% for contamination level of 0.35.