Automated Distribution Network Modeling Using Fine-tuned Large Language Models

New network infrastructure presents a risk to wider power systems, motivating the need to analyze low-voltage distribution network behaviors. Power systems engineers have taken to increasing usage of Large Language Models (LLMs) to conduct detailed and realistic scenarios of network utlisation. However, existing workflows for performing power flow, fault, and reliability analyses, either through general-purpose models or manual coding, remain inefficient, typically requiring intricate topology extraction, detailed parameter definition, and rigorous code validation. To overcome these limitations, this paper introduces a domain-specific fine-tuning strategy for LLMs. This method can generate direct and automated standardized simulation code from geographic data. The proposed method not only automates code creation but also clearly annotates critical parameters. This method also allows engineers to easily modify the code and significantly enhances the efficiency of preliminary modeling processes. Additionally, the performance of various foundational models is compared in this paper. Experimental evaluations indicate that the fine-tuned CodeGemma model achieves superior outcomes, with a syntax accuracy rate of 92.5%. Evaluation also shows a convergence success rate of 31.7% under specified conditions, highlighting the substantial potential of this method to enhance the effectiveness of power system planning.