基于加速深度强化学习的多数据中心算-电资源协同优化调度OA

To improve the participation capability and scheduling efficiency of multiple geographically dispersed data centers in electricity markets,this paper addresses the issues of excessively high dimensionality of action spaces and slow training convergence faced by end-to-end deep reinforcement learning(DRL)when simultaneously considering multi-data-center coordination,quality-of-service constraints,and photovoltaic-energy storage coordination.A data center aggregator(DCA)model is constructed,and a two-layer decision coordinated scheduling framework oriented to computing power-electric power coupling is proposed.By integrating multiple small-scale data centers managed by a single operator,the DCA model optimizes resource management and workload scheduling,and strengthens their competitiveness in dynamic electricity markets.The proposed model adopts a two-layer solution architecture combining DRL and mixed-integer linear programming(MILP).The upper-layer DRL agent handles far-sighted actions that affect future energy states,while the lower-layer MILP solver optimizes myopic actions related to immediate costs,realizing coordinated optimization of long-term strategies and real-time decisions.Simulation results demonstrate that compared with full-action-space DRL,the total cost of the proposed model is reduced by approximately 42.55%,and the training time is shortened to the minute level.In contrast to linear optimization methods,the proposed approach further decreases electricity purchase costs and total comprehensive costs in diverse workload scenarios.The simulation results also verify the impacts of various factors including workload types,energy storage capacity and quality-of-service requirements on scheduling,which reflects the robustness and universality of the proposed method.