| Title |
Local Small LLM-based AI Agent System with Finite State Controller for SCADA Operation Assistance |
| Authors |
김정윤(Jeong-Yoon Kim) ; 이승호(Seung-Ho Lee) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.8.116 |
| Keywords |
District heating SCADA; Small LLM agent; Finite-state controller; LoRA fine-tuning; Operator support |
| Abstract |
This paper proposes a local small-scale LLM-based AI agent for answering district heating operation queries using XAI4HEAT, a dataset built from real SCADA data collected at district heating sites. In this domain, correctly executing procedures and safely abstaining are more important than fluent language generation, because operators must verify point values, period averages, and cross-facility comparisons quickly and reliably. Conventional end-to-end agents depend on free-form generation for both action selection and parameter creation, which often leads to invalid JSON, premature final answers, and excessive clarification in small-model settings. To address this, we reconstruct query?action traces into a Korean query-answering dataset, standardize action names, separate next-action prediction from tool-argument generation, and handle clarification and abstention with fixed templates. We further replace the learned policy in deployment with a finite-state controller (FSC), constraining tool use through explicit state transitions and procedures. Qwen3-0.6B and Llama-3.2-1B-Instruct are fine-tuned locally with LoRA for one epoch as payload generators. In offline replay evaluation on a single RTX 5060 Ti (16GB), the learned pipeline achieved 0/33 (0%) task accomplishment, while FSC+Qwen achieved 8/33 (24.2%), with 14/17 (82.4%) safe non-answers and 1 invalid output. These results show the district heating SCADA assistance system based on operator verification can be implemented and utilized as a practical operational support tool, even with actual industrial site data and environments with limited computational resources. |