A study of nuclear plant heat rate optimization using nonlinear artificial intelligence and linear statistical analysis models

The emphasis of this dissertation is on developing methods by which a combination of multivariate analysis techniques (MAT) and artificial intelligence (Al) procedures can be adapted to on-line, real time monitoring systems for improving nuclear plant thermal efficiency. Present-day first principle models involve performing a heat balance of plant systems and the reactor coolant system. Typical variables involved in the plant data acquisition system usually number one-to-two thousand. The goal of the current work is twofold. First, simulate the heat rate with MAT and Al computer models. The second objective is to selectively reduce the number of predictors to only the most important variables, induce small perturbations around normal operating levels, and evaluate changes in the magnitude of plant efficiency. It is anticipated that making small changes will improve the thermal efficiency of the plant and lead to supplementary cost savings. Conclusions of this report are several. A sensitivity analysis showed the reduction of input variables by dimensionality reduction, i.e., principal component analysis or factor analysis, removes valuable information. Predictors can simply be eliminated from the input space, but dimensionality reduction of the input matrix is not an alternative option. However, perturbation modeling does require data to be standardized and collinear variables removed. Filtering of input data is not recommended except to remove outliers. It's ascertained that perturbation or sensitivity analysis differs from prediction modeling in that two additional requirements are necessary besides the criterion prediction. One is the magnitude of the criterion result given an input perturbation, and second, is the directionality of the model. Directionality is defined as the positive or negative movement of the heat rate (criterion) given a predetermined increase/decrease in predictor value, or input perturbation. While the criterion prediction is still important, it is directionality that determines whether a model is capturing proper changes in system process information. Final results showed that although the secondary-side of a nuclear plant might meet thermodynamic conditions for a steady-flow system, temporal information is needed by the model in order to capture system process information. Modeling of the data is governed by quasi-static range theory, which states data must be closely spaced (in time) and prior temporal information is necessary. The conclusion reached is the perturbation model of a nuclear plant is a time-dependent, dynamic system; all indications as of date show it is also nonlinear. Hence a time-dependent nonlinear modeling method, such as a neural network with time delayed inputs, is needed for sensitivity modeling.