ABSTRACT
This paper presents a method for verifying the accuracy of the measurement of a process variable in power plants. The method is based on the use of a set of process variables dynamically related to the target process variable. The measurements of this set of variables are used as inputs to a neural network to obtain an estimate of the target variable reading. Agreement, within a set tolerance, of the target instrument reading, and the neural network output(s) establishes that the target instrument is operating properly.
The method was applied to the feedwater flow meters in the two feedwater flow loops of the TMI-1 nuclear power plant. These meters are susceptible to frequent fouling that degrades their measurement accuracy. A neural network whose inputs were the readings of a set of reference instruments, was designed to predict both feedwater flow rates simultaneously. A multi-layer feedforward neural network employing the Backpropagation algorithm was used. A normalization procedure based on the mean and standard deviation of each input vector was used to produce a faster training process.
Training and testing were done on data generated from a plant simulation computer program and on TMI-1 operating data from two different operating cycles. The results show that by using proper training data, the neural network can predict the correct flow rates with an absolute relative error of 2%.
