RSynLLM: A risk-aware routing mixture-of-experts large language model for multi-energy load forecasting in large-scale distribution networks

Accurate prediction of multi-energy loads is crucial for ensuring the safe and efficient operation of power distribution networks (PDNs). However, existing methods often overlook the distinct physical characteristics of different energy loads, which makes it challenging to precisely capture their spatio-temporal evolution patterns. Moreover, these methods primarily minimize statistical errors while neglecting the asymmetric impact of prediction errors on downstream decision-making. To address these limitations, we propose a novel model named risk-aware synergy large language model (RSynLLM). This model incorporates a routing mixture-of-experts framework, with tailored load experts designed to capture the unique physical characteristics of various energy loads. A dynamic router adaptively weights these experts according to their contributions. We introduce a risk-aware loss function that asymmetrically penalizes prediction errors while explicitly quantifying tail risks. This drives the large language model (LLM) to actively suppress load underestimation tendencies and mitigate disruption risks, thereby aligning predictions with downstream decision-making objectives. Additionally, customized prompt instructions are designed to guide the LLM in accurately perceiving the spatio-temporal evolution patterns of multi-energy loads. Extensive simulations conducted on a novel large-scale multi-energy PDN dataset, built using real-world data, demonstrate that the proposed RSynLLM consistently achieves state-of-the-art performance in both prediction accuracy and risk awareness. Codes are available at https://github.com/lishijie15/RSynLLM.