Workload Prediction for Proactive Resource Allocation in Large-Scale Cloud-Edge Applications

Proactive resource allocation is critical for ensuring high availability and efficiency in dynamic, heterogeneous cloud-edge environments. This paper presents a practical and extensible framework for workload prediction and autoscaling, with a focus on virtual Content Delivery Networks (vCDNs). We rigorously evaluate a diverse set of forecasting models, including statistical (Seasonal ARIMA), deep learning (LSTM, Bi-LSTM), and online learning (OS-ELM), using extensive real-world workload traces from a production CDN. Each model is assessed in terms of prediction accuracy, computational cost, and responsiveness to support short-term autoscaling decisions. Predicted workloads are fed into a queue-based resource estimation model to drive proactive provisioning of CDN cache servers. Experimental results show that LSTM-based models consistently deliver the highest accuracy, while lightweight methods such as S-ARIMA and OS-ELM offer efficient, low-overhead alternatives suitable for real-time adaptation or fallback. We further explore adaptive model selection strategies and trade-offs between accuracy, cost, and reactivity. The proposed framework enables accurate, low-rejection autoscaling with minimal overhead and is broadly applicable to other cloud-edge services with predictable workload dynamics.