A Dynamic Thresholds based Anomaly Detection Algorithm in Energy Consumption Process of Industrial Equipment

In the context of dual-carbon strategy and dual-energy consumption control targets, energy conservation of industrial equipment is becoming more and more important. However, because of field barrier and experience dependency, energy conservation efficiency is low even with the spreading of IIoT, which leads to low utilization rate of IoT data in turn. Meanwhile, as an important part in energy conservation process, anomaly detection of energy consumption provides the fundamental for realization of energy saving. Data-driven anomaly detection algorithm are mature in academic area while rarely accepted in industrial area, because of interpretability issue of algorithm and complexity properties of industry activities. As a contribution to energy conservation activity in industry, from the view of data-driven anomaly detection of energy consumption of industrial equipment, this paper points out the capabilities that algorithm needs to own (unsupervised, real-time, adaptive, robust, universality), defines volatility, surge as main anomalies for detection, and propose a dynamic threshold based detection algorithm and estimate its feasibility on a real dataset. Experiment result shows average P, R, F1 score 72.1%, 80.1% and 73.1% separately, with remarking 12.1%, 40.1% and 33.1% improvements comparing with baseline model, and 2.37%, 18.9% and 11.2% improvements to DSPOT. Our work in this paper provides a positive effect for improving the efficiency of energy-saving analysis.