Transferable Load Forecasting and Scheduling via Meta-Learned Task Representations

Abstract

This paper addresses the problems of low prediction accuracy and poor generalization of scheduling strategies in multi-tenant cloud environments. It proposes a meta-learning-based method for cross-scenario load prediction and adaptive scheduling. The method consists of two core modules: a task-aware representation embedding mechanism and a meta-optimized scheduling strategy. First, a task-level representation learning model is constructed to extract transferable structural features from historical load sequences. This improves the model's ability to understand heterogeneous tasks. Then, a scheduling policy generator is designed based on a meta-learning framework. It optimizes the initialization of policy parameters through multi-task training, enabling the scheduler to quickly adapt and efficiently allocate resources when new tasks arrive. Comprehensive experiments are conducted on a real-world cloud workload dataset. The results show that the proposed method outperforms existing representative approaches in terms of prediction error, scheduling violation rate, and response latency. It demonstrates good generalization and stability, and effectively enhances resource utilization and service quality in cloud platforms.