UA-FGDRL: An Uncertainty-Aware Federated Multi-Objective Green Deep Reinforcement Learning Framework for Energy-Efficient Cloud Resource Allocation

Cloud resource management remains a big challenge. Cloud workloads are constantly evolving. The demand of the users is frequently unclear. Large cloud systems also require a lot of energy. Data privacy is also a key consideration. These problems make it challenging to manage cloud resources effectively. There are many existing resource allocation methods that are based on heuristic or deep learning. These methods rely on workload prediction. However, prediction errors often occur. The high costs of training these models are also needed. Most of these methods are based on the central collection of data, leading to privacy and scalability issues. To overcome these problems, this paper introduces an Uncertainty-Aware Federated Green Deep Reinforcement Learning (UA-FGDRL) framework. The framework consists of three key elements. The first is a workload prediction model with uncertainty. It helps in handling demand variations. The second one is federated learning. It enables nodes in the cloud to train models without sharing raw data. This is helpful to safeguard information privacy. This also decreases the communication overhead. The third part is a green-aware deep reinforcement learning approach. It aims to increase the efficiency and decrease the energy consumption. The uncertainty-aware prediction module enhances resource provisioning. It decreases the forecasting errors in a dynamic workload. Federated learning allows for decentralized training of cloud nodes. Data is stored at local servers. The green-aware learning agent includes energy and cost factors in the reward function. This contributes to the achievement of balanced resource utilization. Real cloud workload data is used for experiments. The results indicate that the proposed framework can be used to increase the utilization of resources by 15–20%. Up to 17% decrease in SLA violation rates. The energy consumption is reduced by 13-16%. The operational cost is lowered by approximately 15% than recent deep reinforcement learning methods. The model also has a faster convergence. With varying workloads, prediction accuracy is enhanced.