Bridging Continual Learning and Green Cloud Computing: Foundations for Sustainable Time Series Anomaly Detection

Abstract While anomaly detection is essential for cloud computing and predictive maintenance, approaches that bridge continual learning with environmental sustainability remain largely unexplored. In production environments, evolving data distributions cause performance degradation of machine learning models, and naive adaptation may lead to forgetting of past patterns that are likely to reoccur. In this study, we formulate temporal concept-incremental anomaly detection for both univariate and multivariate time series. We introduce time-based and entity-based concept extraction strategies together with Wasserstein-distance diagnostics for inter- and intra-concept distribution drift, and define an evaluation protocol that measures detection quality, knowledge retention, and energy consumption. We further propose Selective Temporal Replay (STR), a replay strategy for autoencoder-based continual anomaly detection that combines diversity-based initialization, novelty-gated admission, retention-aware replacement, and age-weighted preservation of stored sequences. The empirical study comprises more than 5,400 experiments across four datasets, nine autoencoder architectures, and ten continual learning strategies. On Yahoo! A1, STR obtains the highest ROC-AUC in seven of the nine architectures and is the only strategy with positive backward transfer on all four datasets; under the smallest memory budgets, it obtains the highest ROC-AUC among replay methods on Yahoo! A1 and SMD. The sustainability analysis identifies when selective replay justifies its additional energy cost and when regularization-based continual learning strategies are preferable.