Adaptive Load Balancing in Multi-Cloud Systems Using ResFedLB for Load Balancing and RL-COSCoati with Fuzzy Fault Management

Abstract

The growing use of cloud services in the contemporary society has put unprecedented pressure on scalability, responsiveness, and fault resilience, especially in multi-cloud environments that combine heterogeneous resources across providers. Such systems have been difficult to balance their loads effectively because of workload variability, varying Service Level Agreements (SLAs), and the necessity to allocate tasks efficiently with energy constraints of Ultra-Reliable Low-Latency Communication (URLLC). Traditional approaches, such as heuristic schedulers, metaheuristic optimizers, and Reinforcement Learning (RL)-based solutions, have provided partial solutions but still have limitations in the form of slower convergence, high task latency, high migration overhead, and poor fault-tolerant behavior. In order to address these problems, a new framework is presented. The model incorporates four significant innovations. To facilitate privacy-preserving workload forecasting, Resource Aware Federated Learning to Load Balancing (ResFedLB) is first used. Second, CNN -LSTM -Attention with Forecasting Transformer Network (CLAFT-Net) is the specialized local training module that is used to learn temporal-spatial workload dependencies. Third, Reinforcement Learning with Coordination and Opposition Strategy in Coati (RL-COSCoati) is an adaptive and low-latency task scheduling. Lastly, CloudGuard is a fuzzy inference-based module, which guarantees proactive fault detection and resilient VM management. The framework is proven to be effective in experimental assessments with a forecasting accuracy of 99.38 and an F1-score of 98.32, and task allocation performance of 97% Central Processing Unit (CPU) utilization and 98.20% task success rate. These findings prove the hypothesis that the suggested solution provides a scalable, energy efficient, and fault-tolerant solution to managing dynamic workloads in multi-cloud ecosystems.