NARX Structure: One-Step/Multi-Step Ahead Prediction

After model training with the NARX Structure method, two scenarios for the application of the model in the prediction can be distinguished.

One-Step Ahead Prediction

In the case of a one-step ahead prediction, the past system outputs are known and given by actual measurements, e.g. through sensors. The model has to predict just the upcoming time step.

Multi-Step Ahead Prediction

In the multi-step ahead prediction, the past system output values are replaced by the model’s predictions. This corresponds to an offline simulation and is the standard use case, where the user wants the model response to a given set of input variations.

Compared to the one-step ahead implementation, the multi-step case usually results in a worse model quality due to the accumulation of prediction errors.

 

Fig. 26: Model structure for nonlinear autoregression with exogenous inputs (NARX)