ENHANCING WORKLOAD PREDICTION ACCURACY IN SERVERLESS EDGE COMPUTING USING A LEARNING-DRIVEN HYBRID ADAPTIVE MODEL

Abstract

Serverless edge computing has emerged as a critical infrastructure for next-generation distributed systems; however, it faces fundamental challenges in resource management due to highly dynamic workloads, bursty request patterns, and stringent edge hardware constraints. Existing approaches—either purely relying on LSTM-based time-series forecasting or adopting reactive Deep Reinforcement Learning (Deep RL) strategies—fail to achieve an optimal trade-off between Quality of Service (QoS) and resource efficiency. To address this limitation, this paper proposes HybridEdge, a hybrid and adaptive framework that tightly integrates multi-step workload forecasting with adaptive reinforcement learning based on Proximal Policy Optimization (PPO) within a unified control architecture. HybridEdge explicitly incorporates uncertainty quantification via Monte Carlo Dropout and online concept drift detection using the Page–Hinkley test, enabling self-regulating and risk-aware resource provisioning policies. Extensive evaluations conducted on real-world Azure Functions traces demonstrate that HybridEdge achieves a 24.4% reduction in over-provisioning compared to DRL-Only and a 66% reduction in under-provisioning relative to LSTM-Only, while simultaneously maintaining 91.42% user satisfaction and an F1-score of 98.76%. These results confirm that integrating uncertainty-aware multi-step forecasting with adaptive reinforcement learning provides a scalable and effective solution for intelligent resource orchestration in Serverless Edge–Cloud environments, paving the way for the development of autonomous and self-adaptive systems in this domain.