Reinforcement Learning-Based Task Scheduling Under Energy Budget Constraints in Edge-Cloud Systems

Abstract

This paper addresses the energy constraint challenge in task scheduling within edge-cloud collaborative architectures by proposing a scheduling method named EA-BCRL (Energy-Aware Budget-Constrained Reinforcement Learning). The method integrates reinforcement learning with an energy-aware mechanism. It constructs an energy budget modeling module to dynamically estimate the current energy state and budget utilization of the system, enabling real-time modeling and feedback of budget constraints. A budget-guided reward shaping mechanism is introduced to embed energy constraint signals into the policy optimization process, guiding the reinforcement learning agent to avoid budget violations while ensuring task performance. The overall framework achieves coordinated optimization of energy modeling, policy updating, and budget control during task offloading and resource scheduling. Experiments are conducted on a real-world dataset to validate the method at scale, with evaluation metrics including task success rate, average energy consumption, budget violation rate, and average latency. Comparative results show that the proposed method outperforms existing mainstream scheduling approaches across multiple key performance indicators. Ablation studies are further carried out to analyze the independent contributions of each module. Multi-dimensional sensitivity tests are also performed on learning rate, task scale, data disturbance, and task heterogeneity. These results confirm the stability and adaptability of the method. Overall, the findings demonstrate that the proposed approach enables efficient and stable energy-constrained scheduling under complex resource environments and diverse workload conditions.