Towards novelty detection in electronic devices based on their energy consumption

Abstract

Electricity in Brazil is mostly generated by hydroelectric plants that depend on the volume of water in their reservoirs. Due to the fact that rainfall is dramatically decreasing year by year, alternative methods, much more expensive, are often required to supply the energy demand. The increasing number of electronic devices, overconsumption, and energy wasting are also contributing to the problem. There are many ways for wasting energy, often as a result of malfunction devices or human faults. In this way, to assist consumers to save energy and repair a possibly damaged equipment, we propose a system to monitor the energy consumption of electronic devices in order to automatically detect novelties and send alerts. For this, we have evaluated the performance of established machine learning methods, such as Sliding Window, Exponentially Weighted Moving Averages, Clustering, Average consumption by Cycle and Stage, Gauss Distribution, and Artificial Neural Networks. The results show that such methods are very efficient in real-time novelties detection, since they have presented a balanced performance with a high novelty detection rate and low false alarm rate.