Long-range nonlinear prediction: a case study

Long range nonlinear prediction problems can hardly be tackled with classical prediction error based identification methods, which often obtain redundant models with unsatisfactory and possibly unstable performance in simulation. A novel identification algorithm is developed for polynomial NARX models, which combines a model selection procedure based on the minimization of the simulation error and a pruning mechanism for the elimination of redundant terms. The effectiveness of the algorithm is evaluated on a benchmark application example, the long range prediction of the radial crest displacement in the Schlegeis Arch Dam.