Real-time anomaly detection in seasonal time series with conditional variational autoencoder

Real-time anomaly detection in high-frequency seasonal time series is commonly addressed using prediction-based methods, which require waiting for new values to perform subsequent predictions and demand continuous processing over time. This work introduces a novel framework for real-time anomaly detection in seasonal time series, with a practical implementation using Conditional Variational Autoencoders based on Multilayer Perceptrons. Our approach eliminates the need for historical time series data at inference time, instead generating a one-shot long-term expected time series that enables immediate evaluation of streaming data with minimal computational resources. Empirical evaluations on real-world seasonal time series demonstrate that the proposed approach achieves state-of-the-art performance compared in both semi-supervised and unsupervised settings. The framework provides computational efficiency and low energy consumption, making it suitable for deployment in commodity hardware and offline environments.