Development and application of multi-layer neural networks for estimation of power plant variables

For effective control strategies in process industry systems, it is necessary to validate plant sensors and to monitor a multitude of variables. Signal validation and process monitoring problems in many cases require the prediction of one or more process variables in a system. The feasibility of using neural networks to characterize process variables as a function of other related variables is studied. These neural networks are implemented as signal estimation models. This is a departure from the traditional discrete classification problem. The Backpropagation Network (BPN) is used to generate "network models" of signals for both a Pressurized Water Reactor (PWR) and the Experimental Breeder Reactor-II (EBR-II). Several innovations are made in the basic algorithm to accelerate the convergence and improve accuracy. The most significant among these is the progressive adjustment of the sigmoidal threshold function. An interactive BPN algorithm is implemented in a VAX workstation and includes the changes made to the basic algorithm. The following four problems are studied:

1. Multiple-input single-output heteroassociative networks for signal validation for distributed sensor arrays.

2. Multiple-input multiple-output autoassociative networks for plant-wide monitoring of a set of process variables for diagnostics.

3. Identification of plant states, start-up, normal operation, and shut-down, using PWR data.

4. Estimation of the moderator temperature coefficient reactivity, Qc, using PWR data. The dynamic form of the Backpropagation Network is studied for estimating the reactor power of a PWR. The purpose of this study is to take into account any delay between the responses of the plant and the sensor output. Inclusion of several previous readings and using them as inputs to the network has improved the network accuracy, even though it has increased the training time. An empirical formula is developed for estimating the optimum size of the hidden layer nodes. This criterion is closely related to Shannon's information.