Abstract
Accurate prediction of refrigerant deficiency in consumer air conditioning (AC) systems is critical for optimizing energy efficiency and operational stability. However, existing data-driven models often suffer from significant performance degradation due to domain shift across different AC types and a lack of explanatory transparency. To address these challenges, we propose AC-RPX (AC-Refrigerant Prediction eXplainable AI), a unified framework that integrates a Domain Encoder augmented with domain-specific tokens and a Large Language Model (LLM) adapted via Low-Rank Adaptation (LoRA). The Domain Encoder aligns sensor distributions across six consumer AC types to enable domain-robust refrigerant level prediction. Simultaneously, the LoRA-tuned LLM, trained on 280,000 expert-aligned sensor-reasoning pairs, generates case-specific chain-of-thought (CoT) explanations that explicitly link abnormal sensor patterns to the predicted refrigerant charge state. Validation on six AC sensor datasets demonstrates that AC-RPX achieves state-of-the-art accuracy and F1 scores, significantly outperforming conventional deep learning and domain adaptation baselines. Moreover, by providing intuitive natural language explanations, AC-RPX improves predictive performance over conventional rule-based heuristic methods while providing clearer diagnostic explanations on real-world consumer data. These results establish AC-RPX as a practical solution for automated refrigerant monitoring and decision support in real-world maintenance services.