• 목록

[Abstract]

Thermal power plants consist of complex and nonstationary processes, where unexpected operational interruptions can result in substantial losses, necessitating reliable and robust early fault detection for effective monitoring. However, single-model approaches face inherent limitations in such environments. This study proposes an Online Adaptive Conformal Prediction (OACP) framework based on hybrid residuals. The framework employs rolling quantile estimation and Mondrian conditioning to address temporal variability and uncertainty arising from changing operational regimes. We evaluated this model using actual data from Korea South-East Power Co., Ltd. (KOEN) supplemented with synthetic fault scenarios including spikes, level shifts, and sensor noise. Experimental results demonstrate that OACP maintains coverage close to the target of 0.95, achieving 0.94 coverage with a time-to-detection of 55 seconds, suitable for practical field application. Through these findings, this research validates the applicability of the proposed OACP framework in highly variable real power plant environments.