LLM-coordination in auto-bidding of frequency regulation: Cross-attention distributional reinforcement agentic learning

Battery energy storage systems (BESSs) involve substantial capital investment, making profitability a critical concern. In Australia’s National Electricity Market (NEM), Frequency Control Ancillary Services (FCAS) markets have become a primary revenue source for BESSs. However, the high volatility and complex dynamics of FCAS markets make it highly challenging to capture potential market opportunities and formulate profitable bidding strategies. To address these challenges, this paper develops a large language model (LLM)-coordinated auto-bidding system for BESS in the joint energy and FCAS markets. The system introduces an LLM-agentic coordination framework to support automated and interpretable bidding through a multi-agent workflow and theory of mind (ToM) reasoning-based interaction. To recognize critical market patterns, a cross-attention architecture is proposed to jointly model trend and periodic features for enhanced representation of market dynamics. Then, a Soft Actor-Critic (SAC)-based distributional deep reinforcement learning (dDRL) algorithm is developed to optimize bidding strategies under epistemic and aleatoric uncertainties inherent in FCAS markets by continuously learning return distributions. The system is validated using actual South Australian market data, and the experimental results demonstrate that the proposed approach consistently outperforms traditional predict-then-optimize (PTO) methods and other baseline DRL algorithms in terms of profitability.