Multi-Objective Optimal Scheduling of Data Centres Based on NSGA-II and Reinforcement Learning
Buqiao Deng, Qing Ai · 2025
With the proposal of 'dual-carbon' goal, data centres need to solve the problems of high energy consumption and high carbon emission while improving computing power support capability. In this paper, a multi-objective intelligent scheduling optimisation framework is constructed around the goal of green computing, jointly considering multi-dimensional indicators such as task latency, energy utilisation, PUE and carbon emission. On the basis of constructing a simulation platform and a real scheduling system, a collaborative regulation algorithm integrating deep reinforcement learning and multi-objective evolutionary strategy is designed. The experiments use real data centre operation data and public benchmark datasets to conduct comparison tests under different load scenarios. The results show that this paper's method outperforms traditional genetic algorithms and single-objective reinforcement learning methods in terms of PUE reduction, task latency reduction and carbon emission control, and has good convergence, stability and adaptability.