Adaptive Resource Scheduling in Multi-Cloud Computing Using Recurrent Neural Forecasting and Memory-Based Metaheuristic Optimization

Abstract Efficient and intelligent task scheduling in heterogeneous multi-cloud environments remains a complex challenge due to conflicting objectives such as energy consumption, delay minimization, service-level agreement (SLA) compliance, and host utilization. This paper proposes a hybrid framework that integrates Long Short-Term Memory (LSTM) networks for temporal workload forecasting with Elephant Herding Optimization with Memory (EHOM) for multi-objective task-to-host allocation. The framework is implemented as containerized microservices on an OpenStack Yoga private cloud with Prometheus telemetry, Kafka streaming, TensorFlow Serving for LSTM-based forecasting, and a Python-based EHOM optimizer orchestrated through OpenStack. Performance is benchmarked against Trust-Aware Spring Swarm Optimization (TSSO), Auto Clipped Double Deep Q-Learning (Auto-CDDQL), Self-Adaptive Flower Pollination-based RSA (SA-FPRSA), Elephant Herding Lion Optimizer (EHLO), and a Markov-based scheduler. Experimental results across workloads of 200–1000 tasks show that the proposed method reduces total energy consumption by 12–22%, decreases normalized delay by 8–15%, and improves deadline satisfaction ratio (DSR) by 2.5–5.3 percentage oints, while consistently maintaining availability and reliability above 97%. These improvements confirm the robustness, scalability, and real-time applicability of the proposed framework for SLA-sensitive multi-cloud environments. The system links LSTM forecasts with an EHOM-based allocator in a closed loop.