Transfer learning for IT operations: A comparison of statistical and neural time-series models for CPU utilization forecasting

Résumé fourni par la source

AI-driven IT operations (AIOps) require forecasting models that combine predictive accuracy with low computational overhead across heterogeneous cloud infrastructures. However, existing studies rarely compare statistical, recurrent, and Transformer-based forecasting models under transfer learning (TL) and deployment-oriented constraints. This paper presents a comparative evaluation of statistical and neural time-series forecasting models for CPU utilization prediction in virtual machine and containerized environments. We analyze ARIMA, RNN, LSTM, Bidirectional LSTM, Informer, Temporal Fusion Transformer (TFT), and a VM-Temporal Transformer using TL between the Microsoft Azure Trace and Alibaba Cloud Trace datasets. Models are pre-trained on Azure virtual machine workloads and fine-tuned on Alibaba container traces to assess portability across heterogeneous infrastructures. Results show that attention-based architectures achieve competitive forecasting accuracy, with Informer reaching a minimum MAE of 0.0267. However, these gains require higher inference latency and larger memory footprints. Conversely, LSTM-based architectures offer the best trade-off between accuracy, inference speed, and storage efficiency, achieving sub-millisecond inference latency (0.2361 ms) and a compact 201 KB model size while maintaining robust performance across multiple forecasting horizons. The study demonstrates that TL enables effective knowledge transfer between VM and container-based environments, improving the portability of workload forecasting models for cloud and edge infrastructures. Beyond accuracy, the paper provides a deployment-oriented analysis of latency, storage footprint, and adaptability, offering practical guidelines for selecting forecasting architectures in resource-constrained AIOps scenarios supporting resilient and sustainable digital infrastructures.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.