Mitigating Power Fluctuations of Large Language Model Training via Hybrid Energy Storage Systems in Renewable-Powered Computing Centers

As the demand for computing power surges, renewable-powered computing centers (RPCCs) are being deployed at an accelerating pace. During large language model (LLM) training, RPCCs exhibit highly fluctuating, broadband load characteristics that can excite grid resonances and thereby pose significant threats to the secure and stable operation of emerging power systems. To address this challenge, this paper proposes a hybrid energy storage system (HESS) planning approach tailored to mitigate the power fluctuations of RPCCs. First, a signal decomposition method based on symplectic geometric mode decomposition (SGMD) is developed to capture multi-timescale load power fluctuations, on top of which a HESS power allocation strategy is designed. Then, by explicitly accounting for the power fluctuation characteristics of LLM training, an energy storage demand assessment method is established and a distributionally robust optimization (DRO) model for HESS planning is formulated. Finally, case studies are conducted to validate the effectiveness of the proposed planning framework and to elucidate the impacts of power fluctuations inherent to different LLM training tasks.