Application of artificial neural network for prediction of operational performance of MNSR

Abstract

The development of models for predicting reactor performance has significance for staff training, research, and education. The present study talks about how to use an artificial neural network to create a model that replicates the changing behavior of the Isfahan MNSR. There are sufficient data points on a variety of variables corresponding to the long-term operation of this reactor to train neural networks. Time, power, and the initial temperature are chosen as input variables, while the temperature of the core's inlet, the core's outlet, the pool, and the control rod position are chosen as output variables. Two distinct artificial neural networks (ANNs) that were trained using various techniques each provided acceptable results after rigorous testing and validation. The first artificial neural network (ANN1) was trained using different data. Case (I): operations with fixed power, and Case (II): operations involving power level up or down. The second ANN was trained using all operational data. The neural network for fixed values has very high accuracy for fixed inputs, according to the results. The next case (II), however, can deliver a result appropriate for a range of power levels. The second ANN mixes the performance of other networks. ANN1-case I has the greatest R-squared value (0.98977), and the results indicated that ANN1-case II and ANN2 are 5.8% and 4.8% less accurate, respectively.