A Deep Reinforcement Learning Dispatch Method for Integrated Energy Systems Considering Data Center Computing-Power

Abstract

With the rapid development of the digital economy, the construction scale of data centers continues to expand. Given the substantial heat generated by their high electricity consumption, data centers have become critical components of integrated energy systems, exhibiting computing–electricity–heat multi-energy coupling characteristics. This coupling necessitates the coordinated dispatch of multiple devices and the simultaneous optimization of multiple objectives. To address this challenge, this paper proposes the CA-CPEn-TD3 algorithm to solve the optimal dispatch problem of a data-center-based integrated energy system. First, an integrated energy system model is established, in which the computing-load-shifting capability of the computing center is considered. By jointly accounting for computing, electricity, heat, and environmental costs, a reinforcement learning optimization objective is formulated. On this basis, building upon the TD3 framework, a constraint-aware action projection mechanism is introduced to effectively prevent violations of hard constraints and to yield more feasible and practical dispatch strategies. In addition, curriculum learning, parallel exploration, and an ensemble decision-making mechanism are incorporated to enhance training stability. Case studies verify that the proposed algorithm effectively provides deployable dispatch strategies and achieves economically efficient and environmentally friendly scheduling amid the continuous expansion of data centers.