Gain confidence of a neural network used for model based control

In a model based control approach that uses a neural network, it is critical that the sign of the gain associated with each input-output pair of the neural model be correct. If the sign of the gain is not correct, it is possible that the controller may move an input in the wrong direction resulting in poor performance. In this paper, a novel algorithm for computing the probability of determining the correct signs of the gains of a neural network model is presented. The results of the algorithm can be used to determine whether sufficient data has been collected for training the model. If sufficient data has not been collected, either additional data can be collected or the inputs with low gain confidence can be eliminated from the model. The algorithm was used to train a set of neural network models that predicted changes in the emissions of nitrogen oxides, NO_x, and carbon monoxide, CO, from a 730 megawatt coal-fired boiler located at Xcel's Sherco Plant in Becker, Minnesota. The trained neural network models were used in a model based control scheme to reduce NO_x emissions from the boiler by 10%.