A Two-Stage Energy Anomaly Detection for Edge-based Building Internet of Things (BIoT) Applications

Abstract

The Building Internet of Energy (BIoE) is quite promising for curtailing energy consumption, reducing costs, and promoting building transformation. Integrating Artificial Intelligence into the BIoE is essential for big data analysis and intelligent decision-making. Typically, using deep learning to predict energy consumption and detect abnormal energy usage is gaining growing interest in BIoE. However, most models use supervised learning and, thus, require data annotation for model training. This is a tough and costly task, which is often performed by experts. This paper proposes an intelligent Anomaly Detection of Energy Consumption approach using an improved two-stage, hybrid supervised-unsupervised learning process. Specifically, to detect abnormal energy consumption, our methodology identifies the anomalies by analyzing the shape of daily energy usage curves themselves instead of the consumption values. Thus, energy consumption profiles were split into weekday and weekend classes. Then, eXtreme Gradient Boosting (XGBoost) is adopted to build a regression model that enables labeling consumption anomalies of the weekdays class using a rule-based algorithm and residuals. Following, unsupervised anomaly detection is conducted using an Isolation Forest algorithm. Next, the abnormalities detected from the two stages are combined. The empirical evaluation of the proposed scheme illustrates promising anomaly detection accuracy, which has reached 95.93%.