Loss of coolant accident tracking algorithm development using dynamic bayesian network

Abstract

In a pressurized water reactor, there are a lot of reactor instruments to inform the operator of changes in the operating conditions. However, until now, the judgement of the root cause of the changes has been mostly made by the operators, which sometimes can be flawed depending on a variety of circumstances. The probability of making the wrong judgement would decrease if there is a more objective method to aid the operator by interpreting the signals from the instrument with an intelligent cognition. Therefore, an artificial cognitive system that can recognize the nuclear reactor state is suggested in this thesis. Two reactor operating states are considered: Normal state and loss of coolant accident (LOCA) state. LOCA can be subdivided into small break, middle break, and large break LOCAs according to the size of rupture area. Also, it can be divided into hot leg and cold leg LOCAs depending on the break location. The artificial cognitive system predicts the state of the reactor in real time, and it determines the type of LOCA if it predicts that accident occurs. The system uses only reactor protection system (RPS) monitoring parameters among various measurements. The proposed system is composed of two dynamic Bayesian models: the Reactor State Determination Model, and the Accident Type Categorization Model. Two models use different directed acyclic graphs, distributions, and assumptions. The Accident Type Categorization Model uses Extreme Value distribution according to the Anderson-Darling test, and the Reactor State Determination Model uses discrete distribution.

The performance of the trained system was evaluated using 360 randomly sampled data. When an accident occurs, the model took 0.3 seconds to recognize it. In particular, it can be seen that all the accidents are identified within 1 second, and the artificial cognitive system performance is good for recognizing normal versus accident states. The average accuracy of the system is about 88% in case of classifying LOCA type. Also, the average accuracy increased by 4% when the steam generator pressure was excluded from the cognitive system. Especially, when many observation parameters are used, the accuracy decreases if the steam generator pressure is included. The convergence of searching parameter to the accuracy and the accuracy of system is highly sensitive to the measurement are identified areas for the future improvement.