Optimizing Nuclear Reactor Control Using Soft Computing Techniques

This report presents the results of an investigation of different soft computing techniques for application in nuclear reactor control. All research and tests have been carried out using a computer simulation of a demonstration model, consisting of an emptying water tank which is refilled by two controlled flows. This demo provides a safe and representative test bed, because despite of it’s simplicity it provides a highly non-linear control problem. Starting with a standard fuzzy logic controller, different techniques have been applied to improve the performance and robustness and to generalize the controller design process. Rule base adaptation using only two simple guiding rules and membership optimization using genetic algorithms have resulted in a promising method to design high-performance controllers for industrial applications. To enable off-line adaptation and optimization, which would be needed due to safety regulations, plant modeling using neural networks has been investigated. Although this technique matches best with the characteristics of the other methods applied (problem independence, no mathematical formulation is needed, high robustness), this appears to be very difficult because of the inaccuracy inherent to neural networks.