Hybrid Energy-Efficient Reinforcement And Optimization with Carbon-Intensity-Based Container Scheduling

Abstract

The rapid growth of containerized cloud applications has significantly increased energy consumption and CO₂ emissions in large-scale data centers, creating major sustainability challenges. This study proposes HERO-CS 2.0 (Hybrid Energy-Efficient Reinforcement and Optimization with Carbon-Intensity-Based Container Scheduling), a scalable carbon-aware scheduling framework that integrates Power-Aware Best Fit Decreasing (PABFD), Dynamic Voltage and Frequency Scaling (DVFS), and Deep Q-Network (DQN)-based reinforcement learning. The framework combines heuristic optimization, hardware-aware power management, adaptive learning, and real-time carbon-intensity-aware scheduling using external energy-monitoring APIs to identify low-carbon computing nodes. Implemented using CloudSim Plus, Kubernetes, Java, and TensorFlow/PyTorch, the proposed model enhances energy efficiency and sustainability in cloud environments. Experimental results demonstrate significant improvements, including energy consumption reduction from 710 kWh to 470 kWh (approximately 33% savings), SLA violation reduction from 4.2% to 2.0%, and execution time improvement from 25.5 seconds to 15.3 seconds. Additionally, carbon intensity impact decreases from 0.81 kg to 0.45 kg. The framework achieves 92% resource allocation accuracy, 95% QoS satisfaction, and 91% host consolidation efficiency. HERO-CS 2.0 is scalable and suitable for deployment in distributed cloud infrastructures, supporting the development of sustainable and green cloud computing environments.