Research on Adaptive Updating Method of Nuclear Power Plant Transient Models Based on Concept Drift

Abstract The transient in Nuclear Power Plant (NPP) can be considered as the process of the system transitioning from one condition to another. By monitoring the time series data during the operation of NPP and utilizing data-driven machine learning approach to build classification and prediction models to identify and predict the transient, it can provide early warnings of abnormalities, assist operators in making decisions in advance when accidents are in the development stage, thereby improving the safety of the NPP. Traditional data-driven machine learning method for transient identification in NPP mainly rely on offline static data for training and use online dynamic data for prediction. However, during the operation of NPP, due to the impacts of environment, operation, working conditions, and equipment aging, the time series data will fundamentally change over time, leading to a shift in the joint distribution between the predictive model and the response variable, known as concept drift. This change significantly reduces the predictive capability of the model. To equip machine learning models with the capability for adaptive updating to address this challenge, the Concept Drift Window Retraining (CDWR) method is proposed based on statistical method and sliding window mechanism. This method can also realize incremental learning with small samples, thereby continuously improving the predictive capability online of the machine learning model. The effectiveness of this method has been verified on a modular high-temperature gas-cooled reactor (MHTGR) simulator.