Are Time Series Foundation models good for Energy Anomaly Detection?

Smart energy meters generate large volumes of fine-grained time series data that captures building's energy consumption patterns.This data can be leveraged to detect anomalous energy consumption patterns and reduce energy waste in buildings.Traditional anomaly detection methods for smart meter data rely on statistical and machine learning techniques, which often struggle to model complex temporal patterns, require extensive feature engineering, and have poor scalability.Foundation models for time series, which are pretrained on massive volumes of time series data from diverse domains, have recently emerged as a versatile and scalable tool for time series analysis.They are capable of handling multiple tasks, including anomaly detection, and demonstrating superior performance compared to traditional models in various domains.Despite their potential for cross-domain applications, their applicability to the energy domain and their performance compared to traditional machine learning models remain largely unexplored.Therefore, in this paper, we analyze the applicability and performance of Time Series Foundation Models (TSFMs) for unsupervised energy anomaly detection.Specifically, we compare the performance of two widely used state-of-the-art TSFMs, TimeGPT and MOMENT, against existing anomaly detection techniques in the literature: (a) two statistical methods (Interquartile Range (IQR) and Modified Z-Score (mZ-Score)), (b) two unsupervised machine learning techniques (Isolation Forest and Local Outlier Factor), and (c) a deep learning-based technique, Variational Autoencoder (VAE).Our experimental results, conducted using the LEAD 1.0 dataset, which consists of annotated hourly energy readings of 200 buildings, show that MOMENT outperforms both traditional statistical methods and unsupervised machine learning methods.Our results reveal that fine-tuning of MOMENT marginally improves its performance.VAE trained from scratch surpasses TSFMs in performance despite having a smaller model size.We also analyze the trade-off between performance, scalability, and compute requirements of these models.Our analysis also provides new research directions for using TSFMs in the energy domain.