Anomaly Detection in Power Consumption: A Comprehensive Multi-technique Approach

In a world where energy efficiency and system reliability play a vital role, safeguarding these systems from various anomalies is of utmost significance. This task is demanding and requires a substantial amount of effort and domain knowledge. This study presents a robust approach for anomaly detection in power consumption time series data collected from the SCADA system of Amendis in Tetouan City, Morocco. Investigating both anomaly detection for unlabeled data and the evaluation of unsupervised anomaly detection in power consumption, the research explores the influences of clustering for anomaly detection rate, introduces an effective ensemble approach, and incorporates the excess-mass metric for unsupervised anomaly detection evaluation. The integrated approach involves K-means clustering, ensemble anomaly detection, and LSTM autoencoder. Experiments highlight the effectiveness of the methodology through different evaluation metrics such as accuracy, precision, recall, F1-score, and ROC-AUC. The ensemble method surpasses individual anomaly detection models. The LSTM autoencoder achieves balanced precision, recall, and high accuracy. These results emphasize the potential of the LSTM autoencoder as an effective approach for detecting anomalies in power consumption time series data, contributing to enhanced energy management practices.