Building digital twins (BDTs) can enhance reliability by integrating real-time data with virtual model, yet most studies still treat virtual model development, fault diagnosis, and in-situ calibration as isolated stages, resulting in fragmented workflows, low automation, and limited interpretability. To address these issues, this study introduces a novel LLM-based AI agent method integrating virtual model development, diagnosis, and calibration throughout the entire lifecycle in BDTs. During implementation, domain-specific AI agent is developed by knowledge engineering (prompt information, basic information, tool information, and building information) to avoid hallucinations. Then, different toolkits are used to automatically develop virtual models using the MLP algorithm, detect and diagnose faults through comparing the residual with threshold based on a period of time, and perform Bayesian in-situ calibration to ensure accuracy. Finally, the multi-level interpretable results are generated. A case study on a building HVAC system demonstrates the effectiveness of this method: the virtual models of return temperature of chilled water and supply air temperature achieve high accuracy with RMSE of 0.17 °C and 0.21 °C, faults are diagnosed with 10 consecutive residuals greater than the threshold of 1.16 °C, and calibration successfully reduces RMSE from 1.04 °C to 0.30 °C. Importantly, the LLM-based AI agent not only executes all stages with user prompts but also generates interpretable reports, reducing reliance on expert knowledge. By enabling integrated, automated, and explainable BDTs, this study highlights the methodological novelty of employing LLM-based AI agents to advance intelligent building automation systems.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号