Cloud computing faces growing challenges in energy consumption due to the increasing demand for services and resource usage in data centers. To address this issue, we propose a novel energy-efficient virtual machine (VM) placement strategy that integrates reinforcement learning (Q-learning), a Firefly optimization algorithm, and a VM sensitivity classification model based on random forest and self-organizing map. The proposed method, RLVMP, classifies VMs as sensitive or insensitive and dynamically allocates resources to minimize energy consumption while ensuring compliance with service level agreements (SLAs). Experimental results using the CloudSim simulator, adapted with data from Microsoft Azure, show that our model significantly reduces energy consumption. Specifically, under the lr_1.2_mmt strategy, our model achieves a 5.4% reduction in energy consumption compared to PABFD, 12.8% compared to PSO, and 12% compared to genetic algorithms. Under the iqr_1.5_mc strategy, the reductions are even more significant: 12.11% compared to PABFD, 15.6% compared to PSO, and 18.67% compared to genetic algorithms. Furthermore, our model reduces the number of live migrations, which helps minimize SLA violations. Overall, the combination of Q-learning and the Firefly algorithm enables adaptive, SLA-compliant VM placement with improved energy efficiency.

Cloud data centers provide essential scalable computing resources but often suffer from inefficient resource allocation, resulting in excessive energy consumption and increased carbon emissions. This paper proposes a Q-learning-driven adaptive virtual machine placement strategy that simultaneously optimizes thermal performance and energy efficiency in heterogeneous data centers. The proposed approach explicitly considers the physical location of servers within racks as well as their processor performance-to-power ratio (PPR). By dynamically adjusting CPU utilization thresholds according to the servers’ rack positions, the method ensures that servers operate near their optimal PPR. The algorithm formally classifies servers into “best gear,” “high preferred gear,” and “low preferred gear” states. A reinforcement learning framework based on Q-Learning learns optimal placement policies to minimize energy consumption, maintain stable service-level agreements (SLAs), and reduce the risk of thermal hotspots. Experimental results show that our approach reduces energy usage by 18.43% compared to particle swarm optimization, 20% compared to genetic algorithms, and 13% compared to predictive thermal-aware cloud optimization. Furthermore, it significantly lowers SLA violations and hotspot occurrences. By improving both thermal and energy management, this work contributes to more sustainable, efficient, and environmentally responsible cloud data center operations. Received: 3 June 2025 | Revised: 5 September 2025 | Accepted: 15 September 2025 Conflicts of Interest The authors declare that they have no conflicts of interest in this work. Data Availability Statement The data that support the findings of this study are openly available in the Microsoft Azure 2019 dataset at https://github.com/Azure/AzurePublicDataset, reference number [27]. Author Contribution Statement Abdelhadi Amahrouch: Conceptualization, Methodology, Software, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization. Youssef Saadi: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Resources, Supervision, Project administration. Said El Kafhali: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Resources, Writing – review & editing, Supervision, Project administration.