A domain-enhanced large language model for intelligent risk analysis and autonomous decision support in distribution substations

To address the challenges of insufficient semantic understanding and fragmented multi-modal data in the digital transformation of distribution substations, this paper proposes a Domain-Enhanced Large Language Model (DE-LLM) architecture. This study constructs a four-stage methodological framework: Firstly, domain-adaptive pre-training is conducted based on extensive substation logs and IEC standard documentation to reconstruct the model's semantic space; Second, a multimodal instruction fine-tuning strategy is designed, employing a cross-modal fusion Transformer to achieve semantic alignment between infrared imagery, sensor time-series data, and textual information. Third, an expert feedback reinforcement learning (RLEF) mechanism is introduced, constructing a composite reward function encompassing safety compliance and inference transparency to suppress hallucinations. Fourth, parameter-efficient fine-tuning (LoRA) is implemented to achieve autonomous decision support. Experimental results on a high-fidelity testbed of 15 substations demonstrate that DE-LLM improves risk identification accuracy by 27.9%, reduces alarm response latency by 41.2%, and decreases safety rule violations by 86.4% compared to baselines. These findings confirm that aligning general Large Language Models (LLMs) with industrial physics and safety norms is a viable path toward autonomous substation operations.