13.07.2015 Views

Assessment and Future Directions of Nonlinear Model Predictive ...

Assessment and Future Directions of Nonlinear Model Predictive ...

Assessment and Future Directions of Nonlinear Model Predictive ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Nonlinear</strong> <strong>Model</strong> <strong>Predictive</strong> Control: An Introductory Review 5the current operating vector. These basis functions are chosen to give a valueclose to 1 in regimes where f i is a good approximation to the unknown F <strong>and</strong> avalue close to 0 in other cases.Notice that this technique allows the use <strong>of</strong> a linear predictive controller,avoiding the problems associated to computation time <strong>and</strong> optimality <strong>of</strong> thenonlinear solution. These strategies have been successfully tested on a pH neutralizsationplant (see [45] for details).Neural NetworksThe nonlinear dynamics <strong>of</strong> the process can also be captured by an artificial NeuralNetwork (nn). Neural networks are attractive tools to construct the model<strong>of</strong> nonlinear processes since they have an inherent ability to approximate anynonlinear function to an arbitrary degree <strong>of</strong> accuracy [14]. This, together withthe availability <strong>of</strong> training techniques, has made them very successful in manypredictive control applications <strong>and</strong> commercial products. More details <strong>of</strong> the use<strong>of</strong> nn for control can be found in [32].Neural Networks are usually combined with linear models in practical applications,since they are not able to extrapolate beyond the range <strong>of</strong> their trainingdata set. Based on a model confidence index, the nn is gradually turned <strong>of</strong>f whenits prediction looks unreliable, the predictions relying on the linear part.2.2 State Space <strong>Model</strong>sThe linear state space model can naturally be extended to include nonlineardynamics. The following state space model can be used to describe a nonlinearplantx(t +1)=f(x(t),u(t)) y(t) =g(x(t)) (3)where x(t) is the state vector <strong>and</strong> f <strong>and</strong> g are generic nonlinear functions. Noticethat the same equation can be used for monovariable <strong>and</strong> multivariable processes.Notice also that this model can easily be derived from the differential equationsthat describe the model (if they are known) by converting them into a set <strong>of</strong>first-order equations. <strong>Model</strong> attainment in this case is straightforward but theprocedure is very difficult to obtain from experimental data when no differentialequations are available.This kind <strong>of</strong> model is the most widely extended for nonlinear plants sinceit has given rise to a lot <strong>of</strong> theoretical results: the majority <strong>of</strong> results aboutstability <strong>and</strong> robustness have been developed inside this framework. It is alsoused in commercial tools such as nova nlc or nonlinear pfc.An important class <strong>of</strong> nonlinear state-space models are Piece Wise Affine(pwa) models, which can be used to approximate smooth nonlinear processes<strong>and</strong> are defined by:x k+1 = A i x k + B i u k + f i for: x k ∈X iwhere {X i } s i=1 is a polyhedral partition <strong>of</strong> the state space.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!