Anomalous Electrical Power Consumption Detection in Household Appliances via Micro-Moment Classification

Abstract

The detection of anomalous power consumption is critical for improving energy efficiency, particularly with the increasing demand in buildings. This study explores Convolutional Neural Network-based models by transforming 1-dimensional micro-moment labeled data into 2-dimensional matrices to capture both temporal and spatial consumption patterns. Three architectural variants are investigated: a conventional Deep Convolutional Neural Network (DCNN), a Depthwise Separable Convolutional Neural Network (DS-CNN), and a Depthwise Separable Residual Convolutional Neural Network (DSR-CNN). Unlike earlier studies, this work incorporates hyperparameter tuning, statistical validation, and cross-validation, resulting in the evaluation of over 450 model configurations. The results indicate that while the DCNN consistently achieves the highest accuracy, the DS-CNN achieves comparable performance with significantly reduced parameters and computational cost, making it suitable for real-time and resource-constrained environments. Model complexity analysis and statistical tests confirm the robustness of the findings. Finally, a systematic model selection strategy is presented, identifying the DS-CNN as the most balanced solution for effective and efficient anomaly detection in smart grid applications. © 2020 IEEE.