A study on the optimal fuel loading pattern design in pressurized water reactors using the artificial neural network and the fuzzy rule based system

In pressurized water reactors, the fuel reloading problem has significant meaning in terms of both safety and economic aspects. Therefore the general problem of incore fuel management for a PWR consists of determining the fuel reloading policy for each cycle that minimize unit energy cost under the constraints imposed on various core parameters, e.g., a local power peaking factor and an assembly burnup. This is equivalent that a cycle length is maximized for a given energy cost under the various constraints. Existing optimization methods do not ensure the global optimum solution because of the essential limitation of their searching algorithms. They only find near optimal solutions. To solve this limitation, a hybrid artificial neural network system is developed for the optimal fuel loading pattern design using a fuzzy rule based system and an artificial neural networks. This system finds the patterns that Pmax is lower than the predetermined value and Keff is larger than the reference value. The back-propagation networks are developed to predict PWR core parameters. Reference PWR is an 121-assembly typical PWR. The local power peaking factor and the effective multiplication factor at BOC condition are predicted. To obtain target values of these two parameters, the QCC code are used. Using this code, 1000 training patterns are obtained, randomly. Two networks are constructed, one for Pmax and another for Keff Both of two networks have 21 input layer neurons, 18 output layer neurons, and 120 and 393 hidden layer neurons, respectively. A new learning algorithm is proposed. This is called the advanced adaptive learning algorithm. The weight change step size of this algorithm is optimally varied inversely proportional to the average difference between an actual output value and an ideal target value. This algorithm greatly enhances the convergence speed of a BPN. In case of Pmax prediction, 98% of the untrained patterns are predicted within 6% error, and in case of Keff all of the untrained patterns are predicted within 0.24% error. In same workstation, QCC code takes over ten seconds CPU to calculate output for one loading pattern. On the other hand, BPN predicts after 0.084 sec CPU. Thus BPN predicts about 100 times as fast as the numerical code. The fuzzy rule based system for PWR fuel shuffling is to classify the fuel loading patterns into two groups, good and bad one. This system uses the fuzzy rule for an effective searching. A new method is proposed so that a membership function is automatically adjusted to an ideal one by an adaptive manner. For the first cycle of Kori unit 1 PWR, the fuel loading pattern classification is performed using an ordinary rule based systems and the fuzzy rule based system. The searching space of the fuzzy rule based system is very smaller than that of the ordinary rule based system. Use of the fuzzy membership function removes most of the undesirable patterns. The prototype expert system, the optimal fuel shuffling system (OFSS), is developed. OFSS is composed of five independent processes. The validation problem is solved to demonstrate the capabilities of proposed methods and the developed expert system. The first cycle of Kori unit 1 PWR is selected as the validation problem. From the validation results, it can be concluded that OFSS can find the optimal fuel loading patterns better than the reference one