Abstract
<jats:p>With the rapid advancement of large language models (LLMs), their application in specialized domains has attracted growing attention. However, power-system knowledge is highly specialized, structurally complex, rapidly evolving, and safety-critical, making it difficult for existing LLMs to integrate such knowledge accurately and reliably. To address this issue, this paper proposes an agentic knowledge fusion method for power-system LLMs. Through multi-agent collaboration, the method incorporates planning, task decomposition, reflection, and dynamic routing to retrieve and consistently integrate multi-source domain knowledge. Prompt learning further guides reasoning under explicit knowledge constraints, while local knowledge caching and knowledge-constrained continual pre-training support dynamic knowledge updates and stable capability retention. Experiments on a power-domain benchmark demonstrate significant improvements over existing methods in power question answering and knowledge reasoning.</jats:p>