轻 量 级 大 模 型 驱 动 的 虚 拟 电 厂 人 机 交 互 调 控 方 法

Massive residential users on the load side are characterized by ambiguous information input, frequent demand fluctuations, and highly discrete behaviors. Existing methods struggle to conduct structured characterization, customized modeling and quantitative accounting of the regulation capability of such resources, which has become a key technical bottleneck restricting the practical popularization of virtual power plants among thousands of households. This paper firstly proposes an intelligent modeling method for massive distributed flexible loads based on large models, and designs a full-link human-computer interaction regulation mode for virtual power plants covering user demand input, automatic large model modeling, regulation instruction execution, and human-computer closed-loop feedback. Secondly, a two-stage collaborative large model training method combining supervised learning and reinforcement learning is proposed to realize efficient fine-tuning of low-parameter and lightweight large models. Case study results show that compared with existing methods, the proposed method has remarkable advantages in training efficiency, regulation accuracy and model deployment cost. It provides strong support for fully unleashing the regulation potential of residential users and comprehensively activating flexible regulation resources on the load side.