Predictive Workload Migration Using Federated Learning for Energy-Aware Multi-Site Data Center Management

Abstract

As organizations operate geographically distributed data center portfolios, the opportunity to migrate computational workloads toward sites with lower carbon intensity or surplus renewable energy has become increasingly attractive. However, centralized workload prediction models face challenges related to data privacy, network latency, and heterogeneous infrastructure configurations across sites. This paper proposes FedMigrate, a federated learning-based framework for predictive workload migration that enables collaborative model training across multiple data center sites without sharing raw operational data. Each site trains a local LSTM-Transformer hybrid model on its workload patterns, energy pricing signals, and renewable generation forecasts; a central aggregator coordinates model updates using a weighted federated averaging scheme that accounts for site-specific data quality and infrastructure capacity. We evaluate FedMigrate on a real-world dataset from 6 geographically distributed data centers spanning three continents. Experimental results show that FedMigrate reduces overall carbon emissions by 22% and energy costs by 15% compared to site-local scheduling, while achieving prediction accuracy within 3% of a centralized oracle model that has full data visibility. The framework also demonstrates robustness to non-IID workload distributions and communication-constrained environments