*7.4. Prediction and Simulation Configuration*

In the previous Sections 7.1–7.3, we evaluated the performance for the NN and SM models using the measured output in the model regressor. This configuration can be used for control or estimation applications, to control the plant using the future predicted behaviour. Alternatively, prediction control techniques are desirable. These models are limited to short prediction horizons. On the contrary, if the model is aimed for simulation or pure prediction over long horizons, parallel architecture has to be used as the one in Figure 3. In that case, the model is fed with the previous estimations, and therefore, it can model the whole system's behavior. If the output is disturbed during simulation, the model will not be aligned and will not provide information in this regard.

In order to evaluate the performance of the obtained NN and SM models as predictors, we used Experiments 2 and 3 from Sections 7.2 and 7.3, respectively.

## 7.4.1. Neck Rotation

When these experiments are applied using the parallel architecture, we can see in Figure 20 corresponding to the pitch that both models decrease in performance. However, the dynamics are still well-captured by both models. In the SM case, the limit values are not reached properly with an error of ≈5 deg for the negative picks and ≈3 deg for positive ones. However, the overall dynamics are captured with a fit that marks 77%. For the NN model, the fit marks 58%. As can be seen, there are important dynamic errors in the negative sinusoidal cycle. Regarding the roll, as shown in Figure 21, both models properly capture the dynamics with fits that mark 86.5% for the SM model and 87.5% for the NN model. RLS results are unchanged, since no feedback is used in the regressor.

**Figure 20.** Prediction configuration results for pitch in Test 3.

**Figure 21.** Prediction configuration results for roll in Test 3.
