Existing optimal power flow (OPF) computations typically require manual programming tailored to a fixed operational environment. However, the environment of the distribution network (DN) varies with the frequent changing of modern power systems. This paper proposed an adaptive OPF approach based on large language models (LLMs) to achieve automatic programming of environment changes. First, the approach utilizes publicly available DN datasets and OPF computation scripts, combined with expert knowledge, to construct a supervised fine-tuning (SFT) dataset. Second, the generated dataset is used to perform post-training on low-parameter-scale LLMs, which are suitable for edge-side deployment, using a low-rank adaptation (LoRA) method. Third, to further enhance the accuracy of edge-side LLMs, a feedback self-correction mechanism based on a multi-agent workflow is introduced. Simulation results demonstrate that the proposed approach significantly enhances the capability of edge-side LLMs in solving the OPF problems for the DNs with frequent environment changes.