Risk-aware operation of offshore multi-energy microgrids using large-language-model assisted distributed pareto-optimal reinforcement learning

Given that the operation of offshore multi-energy microgrids (OMEMG) faces various uncertainties, traditional risk-sensitive operation methods often struggle to balance the extreme scenarios and economic return. Particularly, when integrating Conditional Value-at-Risk into the control agent of OMEMG, the scheduling performance under highly volatile or extremely adverse conditions needs to be improved. To provide diversified strategies for agents with different risk preferences, the adaptability and optimization performance should be enhanced. Therefore, a large-language-model assisted distributed Pareto optimal soft actor-critic (LLM-DPSAC) algorithm is proposed. Therein, LLM assists operators in generating scenarios, thus modifying the state transition function of the formulated Markov decision process. By using historical data, LLM can generate more realistic low-probability scenarios, thereby enhancing the model's ability for perception and response in rare situations. Meanwhile, operator-in-the-loop supervision is incorporated to further ensure the reliability and robustness of LLM. Besides, the proposed method accounts for both the average expected value and the entire return distribution, while fully capturing risk-sensitive preferences. The simulation results on an OMEMG system demonstrate that LLM-DPSAC effectively captures the strategy return distributions, balancing the extreme risks and operation cost. Additionally, compared to the SAC algorithm, LLM-DPSAC achieves up to 11.4% higher economic returns and 40% faster convergence across multiple scenarios.