Development of triga mark ii research reactor core monitoring system using adaptive neuro-fuzzy inference system

Abstract

Most of TRIGA research reactors has successfully converted the instrumentation and control (I&C) system from analog-based to digital-based. The digital I&C system is capable to monitor and control variables and parameters as well as to react to the design safety limits and conditions. In this study, the methodology on monitoring three of the core safety-related parameters was developed using the Adaptive Neuro-Fuzzy Inference System (ANFIS) method at Reactor TRIGA PUSPATI (RTP). There were two parts involved which were parameter prediction and deviation calculation. Each parameter was generated with 12 -14 fuzzy inference system (FIS) models according to input-partitioning types. The generated model then underwent the training and testing phases to identify the good fit models which can be calculated based on three statistical calculations which are correlation coefficient (R2), mean absolute error (MAE) and root mean square error (RMSE) to be further validated using a novel dataset. The second part of this study was carried out by constructing the algorithm to calculate the relative error between the predicted parameters and the design safety limit. For validation, the novel RTP dataset was used to select only one good fit model with an optimum input-partitioning method to represent the ANFIS model for parameter prediction in the monitoring system. In fuel temperature reactivity coefficient (FTC) validation, the results show that the Model 12 with fuzzy c-mean and the initial clusters centers of 3 had the lowest MAE and RMSE values which were 0.0110 and 0.1051 respectively however the R2 values are poor; R2 at 0.0795. For the fuel pin power (FPP) parameters at 12 fuel rods radial locations, Model 7 and Model 8 with subtractive clustering as the input-partitioning types and the optimal influenced radius values of 0.40 and 0.45 were selected to represent the FPP parameters at B04 and the rest of the fuel rods. The results show a good accuracy in predicting FPP parameters as the MAE and RMSE were calculated with the lowest values on each of fuel rod. The predicted FPP also shows a strong R2 values of 94% on the average. The validation of the power peaking factor (PPF) at the hot rods determined by the TRIGLAV code also demonstrates a good ANFIS model with 0.45 as the optimal influenced radius value in subtractive clustering input-partitioning types in Model 8. The model results in the lowest MAE and RSME with the R2 values at 0.1844, which is quite low. Although the calculated R2 for FTC and PPF parameters have weak R2 values, this statistical calculation was only used to present the relationship between the actual and prediction output and was not used as the primary model performance evaluation to conclude on the models’ accuracy and capability to predict the parameters. Thus, from these findings, the inclusion of FTC, FPP and PPF with specific optimal input-partitioning type on each ANFIS model can be implemented in the monitoring system for enhancing the reactor safety at TRIGA research reactors.