A Constraint-Based Random Search Algorithm for Optimizing Neural Network Architectures and Ensemble Construction in Detecting Loss of Coolant Accidents in Nuclear Power Plants

One major accident of a nuclear power plant (NPP) is the loss of a coolant accident (LOCA) which is caused by a large break in an inlet header (IH) of a nuclear reactor. This work proposes a constraint-based random search algorithm for optimizing neural network (NN) architectures and ensemble construction in three stages for detecting the break size of an IH of a NPP. In stage one, a number of 2-hidden layer, 3-hidden layer and 4-hiddden layer network architectures are created using a proposed constraint satisfaction algorithm. Then, an optimised 2-hidden layer network, an optimised 3-hidden layer network and an optimised 4-hidden layer network are chosen from these architectures by training and testing them on a transient dataset of IHs and a linear interpolation dataset. In stage two, the optimised 2-hidden layer network, the optimised 3-hidden layer network and the optimised 4-hidden layer network are trained and tested iteratively 200 times on the transient dataset to further improve their performance. In stage three, the optimised 2-hidden layer network, the optimised 3- hidden layer network and the optimised 4-hidden layer network are combined into a neural network ensemble (NNE) using a weighted meaning approach. The results show that the NNE outperformed the individual optimised neural networks in detecting the break size of an IH.