A Hybrid Machine Learning Model for Reliability Evaluation of the Reactor Protection System

Abstract

Reliability evaluation is a vital issue for safety and continuous progress of critical safety system such as Reactor Protection System (RPS) in Nuclear Power Plants (NPPs). In this study, a hybrid reliability evaluation approach of the RPS based on machine learning algorithms is proposed. First, the significant reliability factors in RPS are classified to four sections: Interlocks, over-power DT-shutdown, over-temperature DT-shutdown, and logic circuit reactor shutdown. Each of these sections represents a subsystem of the integrated RPS safety system. Then, a feature selection strategy based on Principal Component Analysis (PCA) is applied to exclude low impact factors. Second, a hybrid Particle Swarm Optimization - Support Vector Regression (PSO-SVR) algorithm is proposed to evaluate the reliability for each subsystem. Third, a hybrid Particle Swarm Optimization – Artificial Neural Network (PSO-ANN) is developed to associate the reliability of the four subsystems with the reliability of the entire RPS. Finally, residual error correction of Markov chain is adapted to enhance the prediction performance of the proposed model. The proposed approach can estimate more accurate prediction error of 0.0892 compared to other the state-of-art models. The RPS unavailability measure using our approach is equal to 1.30E-06 compared to values estimated using other methods.