A Comparative Study of Deep Learning Models for Energy-Aware Workload Prediction in Edge-Cloud Environments

Abstract

The demand for energy-efficient workload management across dispersed edge-cloud infrastructures has increased due to the growing use of edge computing. Conventional reactive scheduling techniques frequently result in higher energy usage and less efficient use of resources. In order to increase energy efficiency, this study proposes a prediction-driven workload management strategy that incorporates deep learning-based forecasting models into an edge-cloud simulation environment. To forecast CPU workload patterns and direct task scheduling choices, four models are used namely, CNN, LSTM, CNN-LSTM, and Transformer. Mobile device tasks are dynamically offloaded to edge or cloud resources in a two-tier edge-cloud architecture with an edge orchestrator. To investigate the relationship between orchestration techniques and prediction accuracy, three task placement policies are used including Utilization based, Network based, and Hybrid. When paired with the Hybrid policy, the Transformer-based model regularly achieves the maximum energy efficiency and the lowest average energy per task. While CNN has the lowest efficiency, the CNN-LSTM model performs competitively. These results demonstrate how well sophisticated deep learning models work when combined with flexible task allocation techniques. The suggested method contributes to the creation of intelligent and sustainable distributed computing environments by offering a scalable and energy-efficient workload management solution for edge-cloud systems.