Using reference trajectories for predicting uncertain systems exemplified for a power plant

This paper presents a method for prediction of uncertain closed loop systems, where the uncertainties are depending on operating points. Such model uncertainties are often present when complicated non-linear systems are predicted. The method uses precomputed mean and variances of the prediction error depending on the operation point given by references and disturbances. These uncertainty models are stored in a model bank, linear interpolation is applied to the elements of the model bank in order to predict uncertainty bounds on the predictions using the statistics of the past prediction uncertainties. It is as well proposed to update the uncertainty prediction models on-line. The potential of the method is illustrated by an example from a coal-fired power plant. This example shows prediction of the uncertainties as a bounded region in which the given system variable can be assumed to be contained in. In the example these successfully bound the system variables while in comparison applying a simple prediction diverges from the system.