Energy-Aware Task Scheduling in Cloud Environments Using Digital Twin Architecture with CNN-LSTM and Multi-Agent Deep Reinforcement Learning

Abstract

The increasing complexity of dispersed cloud networks necessitates smart solutions for energy sustainability and stringent Quality of Service (QoS) maintenance. Existing scheduling approaches, primarily single-agent Deep Reinforcement Learning (DRL), suffer from significant performance and scalability issues in large-scale, dynamic environments. This study proposes a novel, smart, energy-conscious paradigm for task scheduling that integrates a Digital Twin (DT) architecture with Multi-Agent Deep Reinforcement Learning (MADRL). The DT provides a faithful, up-to-the-minute representation of the network, utilizing a hybrid Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) model to forecast future workload demands and server power consumption trends. Using these reliable forecasts, MADRL systems develop a dynamic, decentralized scheduling strategy where each server acts as a cooperative agent. The primary objective is to solve a multi-objective optimization problem that minimizes overall energy consumption (via consolidation and dynamic server shutdown) while adhering to QoS requirements (limiting task latency and increasing throughput). Experimental evaluations, comparing our framework against state-of-the-art single-agent DRL baselines (A3C and DDPG), demonstrate superior performance. Our strategy achieves a significant reduction in response time and operational energy cost, proving the viability of this integrated AI and DT technology for efficiently and scalably managing resource allocation in modern distributed cloud environments.