Simple Self-Tuning Multistep Predictors

Abstract The term multistep prediction is used to indicate the prediction of a stochastical ly disturbed dynamical signal over a longer range of future consecutive sampling instants. This technique is a keystone in some recently developed predictive control strategies with promising application possibilities. The prediction is based on a mathematical model of the dynamical process. When this model is numerically unknown, self-tuning multistep prediction is a valuable alternative. However, its computational burden is rapidly growing with the length of the prediction range. A simple and faster suboptimal multistep predictor is compared to the optimal one. The most attractive features are that the increase in prediction error due to the non-optimality can be computed in real-time from the self-tuner's parameter estimations and that the suboptimal algorithm can be adapted such that this increase does not exceed a prescribed value. Its use for the design of simple self-tuning predictive controllers looks promising.