Predictive feedback control


Published on

Published in: Business
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Predictive feedback control

  1. 1. ISA TRANSACTIONS® ISA Transactions 42 ͑2003͒ 207–226 Predictive feedback control Leonardo L. Giovanini* ´ Electronic and Control Research Group, Universidad Tecnologica Nacional, Regional School at Villa Maria, ´ Av. Universidad 450, 5900 Villa Marıa, Argentina ͑Received 19 July 2001; accepted 29 April 2002͒Abstract In this work a new method for designing predictive controllers for linear single-input/single-output systems ispresented. It uses only one prediction of the process output J time intervals ahead to compute the correspondent futureerror. Then, the predictive feedback controller is defined by introducing a filter which weights the last w predictederrors. In this way, the resulting control action is computed by observing the system future behavior and also byweighting present and past errors. This last feature improves the closed-loop performance to disturbance rejection asshown through simulations of two linear systems and a nonlinear continuous stirred tank reactor. © 2003 ISA—TheInstrumentation, Systems, and Automation Society.Keywords: Predictive control; Model-based control; Feedback output1. Introduction control sequence, and ͑c͒ the first movement of the control sequence is applied. These operations are The use of different kinds of linear models to repeated at time kϩ1.predict the future behavior of the process output The main advantage of MPC is its ability to ad-has stimulated the development of a wide group of dress ͑a͒ long time delay, ͑b͒ inverse response, ͑c͒control methods known as model predictive con- significant nonlinearities, ͑d͒ multivariable inter-trol ͑MPC͒. Many MPC approaches have been actions, and ͑e͒ constraints. The widespread useproposed along the past three decades ͑generalized and success of MPC applications described in thepredictive control ͓1͔, dynamic matrix control ͓2͔, literature ͓4͔ attest to the improved performance ofmodel algorithmic control ͓3͔͒, most of them MPC compared to the classical control algorithmbased on a receding-horizon strategy, i.e., at each for control of difficult process dynamics. How-sampling instant k the following actions are taken: ever, alternative algorithms have been developed͑a͒ the plant model is used to predict the output to address the same problems ͓analytical predictorresponse to a hypothetical set of future control sig- algorithm ͓5͔ ͑APA͒, predictor controller ͓6͔ ͑PC͒,nals, ͑b͒ a function including the cost of future simplified predictive control ͓7͔, variable horizoncontrol actions and future deviations from a refer- predictor ͓8͔ ͑VHP͔͒ because MPC implementa-ence trajectory is optimized to give the best future tions require sophisticated tools that do not allow us to apply predictive control at all levels of con- trol systems. All of these algorithms employ, in different ways, only one prediction of the future *Tel.: 54-353-4537500; fax: 54-353-4535498. E-mail error to compensate time delay and interactions.address: For example, simplified predictive control and PC0019-0578/2003/$ - see front matter © 2003 ISA—The Instrumentation, Systems, and Automation Society.
  2. 2. 208 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226predict the system output at its steady-state value 2. Single-step output predictor͑prediction time is set equal to the convolutionlength͒ and then develop the controller structure. In many predictive control techniques, theOn the other hand, APA predicts the controlled model more frequently used to develop the predic-variable on dead time plus one sample and then tor is the discrete convolution truncated to N termsuses the predicted error as input into the controller. ͓4,9͔. The reason is twofold: ͑a͒ the convolutionFinally, VHP uses the predicted error as the input summation gives the model output explicitly andinto the controller, like APA, but the prediction ͑b͒ the main impulse response coefficients arecan be freely chosen from the hole prediction ho- relatively easy to obtain. In particular, for single-rizon and it does not impose any constraint on the input/single-output ͑SISO͒ systemscontroller that can be used. N In this work a new method for designing a y ͑ J,k ͒ ϭ ͚ ˜ i u ͑ kϩJϪi ͒ , ˆ h JϽN, ͑1͒predictive is presented. The approach is based on iϭ1the use of only one prediction of the system out-put, instead of the complete trajectory: it uses a predicts the output value J sampling intervals ahead, k represents the current time instant tprediction of the process output J time inter-vals ahead to compute the correspondent future ϭkt S ( t S is the sampling interval͒, ˜ i , i herror. The proposed controller, called predictive ϭ1,2, . . . ,N are the impulse response coefficients,feedback controller, uses the last w predicted and u ( kϩJϪi ) , iϭ1,2,...,N is the sequence oferrors instead of using plain feedback errors, as inputs to be considered. However, most frequentlyin classical feedback controllers. Hence the re- y ( J,k ) is not calculated directly from Eq. ͑1͒ but ˆ from a modified expression that includes the pre-sulting control action is computed by observ- diction for the current time y ( 0,k ) . For this, notice ˆing the system future behavior and also by that Eq. ͑1͒ can also be written as a function of theweighting present and past errors. So, this control predicted value for the previous sampling time Jstrategy combines the predictive capacity, which Ϫ1,results in good performance for set-pointchanges and time delay systems, with the classical Nuse of the feedback information which imp- y ͑ J,k ͒ ϭy ͑ JϪ1,k ͒ ϩ ͚ ˜ i ⌬u ͑ kϩJϪi ͒ , ˆ ˆ h iϭ1roves the system performance for disturbance ͑2͒rejection. The organization of the paper is as follow: in where ⌬u ( kϩJϪi ) ϭu ( kϩJϪi ) Ϫu ( kϩJϪiSection 2 the expressions for a general J -step Ϫ1 ) . Then, successive substitutions of y ( J ˆahead output prediction are presented. In Section 3 Ϫ1,k ) by previous predictions givesthe basic formulation for the single-prediction J Ncontroller design is derived. Furthermore, a rela-tionship between the controller parameter and the y ͑ J,k ͒ ϭy ͑ 0,k ͒ ϩ ͚ ˆ ˆ ͚ ˜ i ⌬u ͑ kϩlϪi ͒ . h lϭ1 iϭ1settling time of closed-loop response is estab- ͑3͒lished. In Section 4 the closed-loop stability andperformance of the single-prediction controller are This equation defines a J -step ahead predictor,analyzed. In Section 5 a direct feedback mode is which includes future control actions. Since futureintroduced in order to improve the overall system control actions are unknown, the predictor ͑3͒ isperformance. Besides, the relationship between not realizable. To turn it realizable a statementthe proposed controller and other predictive con- must be made about how the control variable istrol algorithms is established. The closed-loop sta- going to move in the future. For example, the sim-bility and performance of the resulting controller plest rule is to set all of them equal to zero,are analyzed in Section 6. In Section 7 we show ⌬u ͑ kϩ j ͒ ϭ0, ᭙ jϭ0,1,...,J, ͑4͒the results obtained from the application of theproposed algorithm to a nonlinear continuous which implies that the control variable will notstirred tank reactor. Finally, the conclusions are move in future sampling instants. Then, Eq. ͑3͒presented in Section 8. becomes
  3. 3. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 209 J N 3. Single-prediction control y ͑ J,k ͒ ϭy ͑ 0,k ͒ ϩ ͚ ˆ 0 ˆ ͚ ˜ i ⌬u ͑ kϩlϪi ͒ , h lϭ1 iϭlϩ1 Although the idea of using only one prediction ͑5͒ of the system output for controlling the system is not new ͓5– 8͔, all the authors did not use the pre-where the superscript 0 recalls that condition ͑4͒ is diction time J as a tuning parameter. This is theincluded. The new expression ͑5͒ defines a realiz- case of single-prediction controller, whose deriva-able open-loop J -step ahead predictor whose Z tion procedure is quite straightforward from thetransform is given by ͑see Appendix A͒ open-loop predictor. Revising the assumption used to go from Eq. ͑3͒ to Eq. ͑5͒ observe that if just y 0 ͑ J,z ͒ ϭ P ͑ J,z ͒ u ͑ z ͒ , ˆ ͑6͒ the control movement ⌬u ( k ) 0, thenwhere P ( J,z ) is the transfer function of the open- y ͑ J,k ͒ ϭy 0 ͑ J,k ͒ ϩa J ⌬u ͑ k ͒ . ˆ ˆ ˜ ͑9͒loop predictor given by This prediction can be substracted from a refer- N ence variable r ( kϩJ ) to obtain the predicted error P ͑ J,z ͒ ϭa J z Ϫ1 ϩ ˜ ͚ ˜ i z JϪi , h iϭJϩ1 e ͑ J,k ͒ ϭe 0 ͑ J,k ͒ Ϫa J ⌬u ͑ k ͒ . ˆ ˆ ˜ ͑10͒and ˜ J is the Jth coefficient of step response. a The control action can be computed in the similar The prediction y 0 ( J,k ) is updated by adding ˆ way as standard predictive controllers, minimizing the following performance measure, ˜ ͑ J,z ͒ ϭy ͑ J,z ͒ Ϫy ͑ J,z ͒ . d ˆ ͑7͒ f ͑ k ͒ ϭe 2 ͑ J,k ͒ ϩ ␳ ⌬u 2 ͑ k ͒ , ˆ ␳ у0. ͑11͒This term lumps together possible unmeasured Then, the control action that minimizes this per-disturbance and inaccuracies due to plant-model formance index is given bymismatch. Since the future value of ˜ ( J,z ) is not davailable, an estimate is used. In the absence of 1 0 ⌬u ͑ k ͒ ϭ ˆ ͑ J,k ͒ , e ͑12͒any additional knowledge of ˜ ( J,z ) , the predicted d KJdisturbance is assumed to be equal to that esti- where e 0 ( J,k ) ϭr ( kϩJ ) Ϫy 0 ( J,k ) and K J ϭa J ˆ ˆ ˜mated at the current time ˜ ( z ) . A more accurate d ϩ ␳ /a J , and the cost of controlling the system is ˜estimate of ˜ ( J,z ) is possible if the disturbance doutput model and a measure of load disturbance ␳ 2are available. So, we use a similar equation to Eq. f ͑ k ͒ϭ ˆ 0 ͑ J,k ͒ . e ˜ 2ϩ ␳ aJ͑6͒ to predict the future disturbance. The impor-tance of the form of Eq. ͑5͒ comes also from the At this point of the work we are interested onfact that the prediction y 0 ( J,k ) can be updated by ˆ understanding the meaning of prediction time Jthe current output measurement y ( k ) . This is done and its relationship with the closed-loop substituting y ( 0,k ) by y ( k ) , or equivalently the ˆ To do it, we replace ⌬u ( k ) and K J in the closed-correction is implemented by ˜ ( z ) adding to Eq. d loop predicted error ͓Eq. ͑10͔͒, so we can write it͑6͒, in any case we obtain as function of e 0 ( J,k ) and ␳, ˆ ␳ y 0 ͑ J,z ͒ ϭy ͑ z ͒ ϩ ͓ P ͑ J,z ͒ ϪG p ͑ z ͔͒ u ͑ z ͒ , ͑8͒ ˆ ˜ e ͑ J,k ͒ ϭ ˆ ˆ 0 ͑ J,k ͒ . e ˜ 2ϩ ␳ aJ ˜where G p ( z ) is the plant model. This equation Then, defining e ( J,k ) as a fraction of e 0 ( J,k ) , ˆ ˆdefines a realizable corrected J -step ahead predic-tor. Hence there are two type of predictors: one is e ͑ J,k ͒ ϭ ␣ e 0 ͑ J,k ͒ , ˆ ˆ ͉ ␣ ͉ Ͻ1,the open-loop predictor that only depends on thevalues of the past inputs only ͓Eq. ͑6͔͒ and the where ␣ is the remaining error after the controlother is the corrected predictor ͑8͒ which receives action ⌬u ( k ) is applied, the control cost ␳ is re-a correction through the feedback measurement. lated with ␣ through
  4. 4. 210 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 then assumed constant from every sampling in- stant k into the future. In other words, given all the input changes accounted for until the instant k, the single-prediction controller observes the value that would be reached by the system output if no future control action is taken and then u ( k ) is computed such to the performance index ͑11͒ is minimized. Hence if ˜ ( z ) actually remains constant after the d instant k and ␳ ϭ0, then the output reaches the Fig. 1. Structure of the single-prediction controller. reference value J sampling intervals later. Note that controller ͑15͒ is realizable if and only if ˜ J 0. Then, the prediction time J should be a chosen such that ␣ ␳ϭ ˜2. a 1Ϫ ␣ JThis equation means that the closed-loop error Jуint ͩͪtd tS ϩ1,will be bounded by ͉ e ( k 0 ϩJ ) ͉ р ␣ ͉ e 0 ( J,k 0 ) ͉ , ᭙t ˆ where t d is the process time delay. This fact meansϾk 0 ϩJ, where k 0 is time instant where the set- that the open-loop predictor P ( J,z ) compensatespoint changes. Therefore the prediction time J the process time delay. In the following sectionscould be seen as the closed-loop settling time for we assume that ␳ ϭ0 to simplify the expositions,an error ␣ ͉ e ( k 0 ) ͉ . When the control cost ␳ is set however, the results that will be obtained are theto 0, which is equivalent to ␣ ϭ0, the performance same whether the control weight ␳ is set to zero ormeasure ͑11͒ and the predicted closed-loop error not.e ( J,k ) becomes zero, and the control change isˆgiven by 4. Algorithm properties 1 0 4.1. Stability analysis ⌬u ͑ k ͒ ϭ ˆ ͑ J,k ͒ . e ͑13͒ ˜J a To analyze the effect of the prediction time overOn the other hand, when ␳ is set to ϱ ͑which is the closed-loop stability, we substituted the con-equivalent to ␣ ϭ1) no control action is taken and troller ͑15͒ in the closed-loop characteristic equa-e ( J,k ) ϭe 0 ( J,k 0 ) ϭe ( k 0 ) , ᭙kϾk 0 .ˆ ˆ tion to obtain Using Z transform on Eq. ͑12͒ shows that thesingle predictive control algorithm is basically an ͑ 1Ϫz Ϫ1 ͒˜ J ϩ P ͑ J,z ͒ ϩ ͓ Gp ͑ z ͒ ϪG p ͑ z ͔͒ ϭ0. a ˜integral action applied on the predicted error Then, combining this expression with Eqs. ͑A8͒ 1 1 and ͑A2͒, and using the convolution model of the u͑ z ͒ϭ ˆ 0 ͑ J,z ͒ . e ͑14͒ K J 1Ϫz Ϫ1 process, this equation becomes N NCombining Eqs. ͑14͒ and ͑8͒ the result is the con-troller, ˜ Jϩ a ͚ iϭJϩ1 ˜ i z JϪi ϩ ͚ ͑ h i Ϫh i ͒ z Ϫi h iϭ1 ˜ 1 ϱ C͑ z ͒ϭ ͑ 1Ϫz Ϫ1 ͒ K J ϩ P ͑ J,z ͒ ϪG p ͑ z ͒ ˜ . ϩ ͚ iϭNϩ1 h i z Ϫi ϭ0. ͑16͒ ͑15͒ Using a result obtained by Desoer and VidyasagarFig. 1 shows a block diagram of C ( z ) , where it is ͓10͔ for lineal discrete systems, we derive the fol-apparent that it uses the plant model to estimate lowing stability condition ͑see Appendix B͒: ͚ͯ ͯthe output at the present time y ( 0,k ) . This value is ˆ N N ϱthen compared with the actual measurement y ( k )to detect modeling errors and external distur- ͚ iϭJϩ1 ͉˜ i ͉ ϩ ͚ ͉ h i Ϫh i ͉ ϩ h iϭ1 ˜ iϭNϩ1 h i Ͻa J . ˜bances. The global detected disturbance ˜ ( z ) is d ͑17͒
  5. 5. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 211The left side of this equation has, ordered from left Since J stb guarantees the closed-loop stability forto right, the following three terms: ͑a͒ the contri- all the models of W, the open-loop predictorbution of the nominal model, ͑b͒ the additive un- P ( J,z ) can be directly built from the nonlinearcertainty, and ͑c͒ the effect of the truncation model model. This fact improves the accuracy of theerror; all of them are related with the convolution open-loop prediction and the closed-loop perfor-length N. When there is not uncertainty in the mance. The nonlinear predictor can be built fromsystem and if N is large enough to neglect the the nonlinear model employing a numerical inte-truncation error, Eq. ͑17͒ becomes gration scheme or using a local model network N ͓11͔. A final remark is for recalling that the stability ͚ iϭJϩ1 ͉˜ i ͉ Ͻa J stb . h ˜ ͑18͒ condition given by Desoer and Vidyasagar ͓10͔ is a sufficient one. Consequently, stability conditionsThis condition guarantees that—for any time- ͑17͒–͑19͒ become conservative and impose a tooinvariant stable plant—there is always a value of J high lower bound for selecting J. So, always therefor which the closed-loop system is asymptotically are prediction times lower than J stb ; for those thestable. We must note that if the plant has a mono- closed-loop system will be stable. The first J thattone step response, the stability condition ͑18͒ can guarantees the closed-loop stability can be foundbe written as through a direct search, because the solution space ϱ is bounded, 1 1 ˜ J stb Ͼ a ͚ ˜ i ϭ 2 ˜ p, 2 iϭ1 h K 1рJрJ stb . ˜where K p is the process gain. 4.2. Performance analysis Generally, control engineers assume that a fam-ily W of M linear models is capable to capture a The tuning of the single-prediction controllermoderate nonlinearity. Therefore to guarantee the implies a discrete optimization problem having astability of the system we must choose a J such bounded solution spacethat it guarantees the stability of all the plants of J෈N, 0рJрN.W. Then, the robust stability problem becomes theproblem of finding a J such that Eq. ͑17͒ is satis- Hence, independently of the performance indexfied for each model of W. Using global additive being used, the general solution results from a di-uncertainty and choosing N large enough to ne- rect search in the solution space. However, whenglect the truncation error, Eq. ͑17͒ becomes the ISE index is used, the optimal controller re- N N sults from the solution of the following problem: ͚ iϭJϩ1 ͉˜ i ͉ ϩ ͚ max ͉ h li Ϫh i ͉ Ͻa J stb . ͑19͒ h iϭ1 l෈[1,M ] ˜ ˜ min ʈ e ͑ z ͒ ʈ 2 ϭ min ʈ ˜ ͑ z ͓͒ d ͑ z ͒ Ϫr ͑ z ͔͒ ʈ 2 , 2 ␧ Gc(z) Gc(z)Another way to solve the robust stability problem which is equivalent to minimize the sensitivityis to find a J that satisfies simultaneously Eq. ͑18͒ ␧ function ˜ ( z ) over the whole system bandwidth,for all the models of W. In the case of a systemwith monotone response, the stability condition min ʈ ˜ ͑ z ͒ ʈ 2 ϭ min ʈ 1ϪG p ͑ z ͒ Gc ͑ z ͒ ʈ 2 , ␧ ˜͑19͒ can be written as Gc(z) Gc(z) J stb ϭmax͑ J 1 ;J 2 ;...;J M ͒ , and to guarantee the internal stability of the closed-loop system ͓12͔. Under these design con-where J l , lϭ1,2,...,M is the prediction time for ditions the sensitivity function will be minimumthe lth model of the family W. This expression when the open-loop controller Gc ( z ) is given bymeans that we can choose a different predictiontime for each model of W, then we selected the GcϭG p Ϫ1 , ˜ ϩbigger prediction time. At this point, a remark about how to build a ˜ where G p ϩ is the nonminimum phase portion ofpredictor for nonlinear system must be made. the plant ͑right-half plane zeros and time delay͒.
  6. 6. 212 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226For the single-prediction controller, this conditionis equivalent to choose a prediction such that theJth coefficient of the step response ˜ J has the a u ͑ k ͒ ϭu ͑ kϪ1 ͒ ϩ ͚ͭ ͮ V jϭ1 k j e͑ k ͒same sign as the stationary process gain, i.e., V J ISE ϭk i ϩ1, ͑20͒ Ϫ ͚ k j ͓ P ͑ j,z ͒ ϪG p ͑ z ͔͒ u ͑ kϪ1 ͒ , ˜ jϭ1where k i is the number of samples which covers ͑24͒the effect of the minimum phase portion of theplant. where V is the prediction horizon and k j , j When JϭJ ISE , the controller gain ˜ Ϫ1 is the aJ ϭ1,2,...,V is the jth element of the gain vector. Inlargest. Therefore it provides a vigorous control these equations we can see that both predictiveaction and attempts to drive the system output to controllers have a similar structure: the two firstthe reference in J ISE time intervals. In this case, terms, ordered from the left, are a discrete PI con-the single-prediction controller becomes a dead- troller and the last term is a weighing contributionbeat ͑minimum-time͒ controller of the future open-loop deviations at time ( k ϩ j)tS , 1 C͑ z ͒ϭ . ⌬y 0 ͑ j,k ͒ ϭy 0 ͑ j,k ͒ Ϫy ͑ k ͒ , ˆ ˆ ˜ jϭ1,2,...,V,J. Ϫ1 ͑ 1Ϫz ͒˜ J ISE ϩ P ͑ J ISE ,z ͒ ϪG p ͑ z ͒ a ˜ ͑21͒ They only differ in the number of prediction gains employed. So, it is easy to see that we can chooseWhen J is increased the controller gain and the the prediction time J such that both predictiveclosed-loop performance decreases. In case of J controllers, single-prediction and DMC, haveϭN the controller gain is the inverse of process similar performances.gain (˜ J ϭK p ) and the controller drives the sys- a ˜ Example 4.1. To analyze the sensitivity of the pro-tem output to the reference in N time intervals. posed controllers to the parameter J we considerThus only one significant control move is ob- a heat exchanger, whose hot outlet temperature isserved in absence of uncertainty ͑minimum- controlled by manipulating the cold stream flowenergy controller͒ and the single-prediction con- rate, modeled by [13]troller becomes a predictor controller ͓6͔, 35.41 Gp ͑ s ͒ ϭϪ . ͑25͒ 1 ͑ 4.5sϩ1 ͒ 5 C͑ z ͒ϭ . ͑22͒ ˜ N ϪG p ͑ z ͒ a ˜ The discrete model employed to build the single- prediction controller is obtained by assuming a To finish this analysis, we give a heuristic com- zero-order hold at the input of continuous modelparison of the closed-loop performance achieved (25), a sampling time t S ϭ2 and the convolutionby the single-prediction controller with that of a length N was fixed in 50 terms. The predictionMPC controller. To carry out this comparison we time is chosen using the stability condition (18), soanalyze the control actions computed by both con- J must betrollers. They are obtained by replacing the open-loop error by their components, assuming a step Jу12.change in the setpoint, in the controller equations͓4,9͔ and combining them with the predicted out- Fig. 2 shows the responses to a setpoint changeput ͑8͒. The result for the single-prediction con- for different values of J. As it was anticipated,troller is small values of J give more rapid response and require large initial movements in the control vari- u ͑ k ͒ ϭu ͑ kϪ1 ͒ ϩK Ϫ1 e ͑ k ͒ J able (Fig. 3). Observe also that the closed-loop system is stable even for values of J smaller than ϪK Ϫ1 ͓ P ͑ J,z ͒ ϪG p ͑ z ͔͒ u ͑ kϪ1 ͒ , J ˜ the limit provided by Eq. (18). Now, we compare the closed-loop responses ͑23͒ provided by single-prediction controllers with thatand the result for the MPC is provided by a DMC controller. The DMC control-
  7. 7. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 213Fig. 2. Closed-loop responses of the linear system to a step change in the setpoint, showing the effect of the predictiontime J.ler was designed following the tune procedure de- five samples, and the condition number of the con-veloped by Rahul and Cooper [14]. The plant troller c was fixed to 500. The control weight ␭model (25) was approximated through a first- was computed using the following formula [14]:order plus time delay model, Gp ͑ s ͒ ϭϪ35.41 e Ϫ10.1s 12.04sϩ1 , ␭ϭ U c ͩ ␶ 3.5 ϩ2Ϫ tS UϪ1 2 K p ϭ3.73. ˜ ͪwhich was discretized assuming a zero-order hold Fig. 4 shows that the single-prediction control-at the input and a sampling time t S ϭ2. The con- ler can provide a similar performance to that ob-volution length is the same that we used to build tained by the DMC controller. This figure alsothe single-prediction controller (Nϭ50). The pre- shows that J can be selected such that the perfor-diction horizon V was set equal to the convolution mance or the robustness of the system be im-length (VϭN), the control horizon U was fixed to proved, by reducing or increasing J. Fig. 3. Control variable correspondent to the closed-loop responses shown in Fig. 2.
  8. 8. 214 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226Fig. 4. Closed-loop responses of the linear system to a step change in the setpoint, comparing the response of the DMC andsingle-prediction controllers.5. Predictive feedback control the prediction is calculated ͑Fig. 5͒. Hence the control movement ⌬u ( k ) is given by From the analysis of the closed-loop perfor-mance in Section 4.2, it is clear that the single- w ⌬u ͑ k ͒ ϭ ͚ q * e 0 ͑ J,kϪ j ͒ ,prediction controller provides a similar closed- j ˆ ͑27͒loop performance to a MPC controller. This fact jϭ0means that a single-prediction controller shows apoor closed-loop performance when disturbancesand uncertainties are present in the system, spe- where e 0 ( J,kϪ j ) is the J step ahead open-loop ˆcially when they are assumed to be time invariant. error computed at time kϪ j. The predictive feed-This is true even when the underlying system is back controller can be derived from Eq. ͑27͒, re-time invariant ͓15͔. placing the open-loop error e 0 ( J,kϪ j ) by their ˆ A way to solve this problem is to introduce a components and following a similar procedure todirect feedback mode in the computation of the obtain Eq. ͑15͒. The result is the controllercontrol action. This idea can be accomplished byincluding a filter F ( z ) in the single-prediction F͑ z ͒control law ͑14͒. The filter not only includes a C͑ z ͒ϭ ,feedback action in the predictive controller, but ͑ 1Ϫz Ϫ1 ͒ ϩF ͑ z ͓͒ P ͑ J,z ͒ ϪG p ͑ z ͔͒ ˜also introduces a new set of parameters to allow ͑28͒more demanding performances. Since the controllaw ͑14͒ includes an integral mode, F ( z ) can takethe following form: whose structure is shown in Fig. 6. We must note that F ( z ) adds additional degrees of freedom to w w 1 improve the closed-loop performance. This fact F͑ z ͒ϭ ͚ q jz Ϫ jϭ ͚ q *z Ϫ j, j ͑26͒ makes more difficult the controller design since ˜ J jϭ0 a jϭ0 now it is necessary to tune the filter parameterswhere w෈Z is the filter order and q j , j෈ ͓ 0,w ͔ ͑the coefficients q * , jϭ0,1,...,w and the order jare the new controller parameters. Since the con- w) . One way to solve this tuning problem is bytrol law ͑14͒ employs only one prediction of the using a method for a fixed-structure controller likeprocess future behavior, the delay operator z Ϫ j , j those proposed by Abbas and Sawyer ͓16͔ or Har-ϭ0,1,...,w, is applied to the time instant at which ris and Mellichamp ͓17͔.
  9. 9. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 215 Fig. 5. General MPC and predictive feedback setups.5.1. Relationship with other control algorithms w 1 C͑ z ͒ϭ ͚ q *z Ϫ j. ͑ 1Ϫz Ϫ1 ͒ jϭ0 j ͑29͒ It is easy to see that the structure of the predic-tive feedback controller is a generalization of the When the prediction time is set equal to the timeinternal model control parametrization of the feed- delay ( J d t S ϭt d , ) the open-loop predictor P ( J,z )back controllers ͑Fig. 6͒. Depending on the value becomes the system model without time delay, andof the prediction time and the parameters of the the predictive feedback controller ͑28͒ is the Smithcontroller we have the different controllers that predictor of the reduced order controller ͑29͒,would be studied in the specialized literature. When Jϭ1 the open-loop predictor P ( J,z ) be- ˜ F͑ z ͒comes the system model G p ( z ) , and the predic- C͑ z ͒ϭ .tive feedback controller ͑28͒ is a reduced order ͑ 1Ϫz Ϫ1 ͒ ϩF ͑ z ͓͒ 1Ϫz ϪJ d ͔ G p ͑ z ͒ ˜controller ͓6͔, ͑30͒ Fig. 6. Structure of the predictive feedback controller.
  10. 10. 216 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226In this case, the open-loop predictor only compen- This controller is derived from the predictive feed-sates the time delay present in the system. In the back controller ͑28͒ by fixing its parameters tocase of JϭJ d ϩ1 and the parameters of the filter 1 ␤are free to be tuned, the predictive feedback con- wϭ1, q 0 ϭ a Ϫ1 ˜ N , and q 1 ϭ ˜ Ϫ1 . atroller becomes the analytical predictor algorithm 1Ϫ ␤ 1Ϫ ␤ N͓5͔. Observe that varying the parameters ␣ and ␤ gov- When the prediction time is greater than the erns the character of these controllers influencingtime delay ( J d ϽJϽN ) , wϭ0, and q 0 ϭa Ϫ1 the * ˜J the speed of closed-loop response. When the pa-resulting controller is the single-prediction con- rameters are in the lower limit ( ␤ ϭ0 and ␣ ϭ1) ,troller, the controllers ͑33͒ and ͑34͒ become the predictor controller ͑32͒, obtaining the open-loop response. 1 In the other case ( ␤ ϭ1 and ␣ ϭϱ) the controllers C͑ z ͒ϭ , ͑31͒ ˜ J ϩ P ͑ J,z ͒ ϪG p ͑ z ͒ a ˜ ͑34͒ and ͑33͒ become the inverse of the system model, obtaining the minimum time response.which was studied in the previous sections. The Varying the parameters between these limits wecharacter of this controller is governed by the pre- modify the characteristics of the closed-loop re-diction time J, which directly influences the speed sponse, speeding up or slowing down the re-of closed-loop response. For the particular choice sponse, in similar way as the single-predictionof the prediction time JϭN, we can derive a fam- controller with the prediction time J.ily of predictive controllers whose main character- Finally, we can see that predictive feedback con-istic is to obtain a closed-loop response that is at trol has a strong connection and significant differ-least as good as the normalized open-loop re- ence from VHP ͓8͔. The approach employed bysponse. If no other design condition is demanded, both frameworks is similar. They employ one pre-the controller ͑23͒ becomes the predictor control- diction of the system output, which can be freelyler ͓6͔, chosen, and use the predicted error as inputs into the controller. However, VHP by itself is a predic- 1 tive structure, not a controller, that can be added to C͑ z ͒ϭ . ͑32͒ another control algorithm ͑for example PI, PID, or ˜ N ϪG p ͑ z ͒ a ˜ simplified predictive control͒. It is employed to compensates time delay and interactions and pro- a Ϫ1Fixing the parameter of the controller q 0 ϭ ␣˜ N , vide a built-in feedforward scheme. In this way,␣ у1, we obtain simplified model predictive con- VHP is similar to a single-prediction controller.troller ͓7͔, Furthermore, VHP computes the whole trajectory, like standard predictive control, and then selects ␣ one prediction. C͑ z ͒ϭ . ͑33͒ ˜ NϪ ␣ G p͑ z ͒ a ˜The parameter ␣ provides a way to speed up the 6. Predictive feedback propertiesclosed-loop response and build a dead-time com-pensation in the controller, but it does not give 6.1. Stability analysisoffset-free responses in the presence of modelingerrors. To solve this problem, Chawla et al. ͓18͔ Now, we study the effect of the filter and pre-proposed the inclusion of a first-order filter into diction time on the closed-loop stability. Then, wethe control law such that the resultant controller is substitute the predictive feedback controller ͑28͒the conservative model based controller, in the characteristic closed-loop equation, which becomes ͑ 1Ϫ ␤ z Ϫ1 ͒ T ͑ z Ϫ1 ͒ ϭ ͑ 1Ϫz Ϫ1 ͒ ϩF ͑ z Ϫ1 ͒ P ͑ J,z Ϫ1 ͒ C͑ z ͒ϭ , ͑ 1Ϫ ␤ ͒˜ N Ϫ ͑ 1Ϫ ␤ z Ϫ1 ͒ G p ͑ z ͒ a ˜ ϩF ͑ z Ϫ1 ͓͒ Gp ͑ z Ϫ1 ͒ ϪG p ͑ z Ϫ1 ͔͒ . ˜ 0р ␤ Ͻ1. ͑34͒ ͑35͒
  11. 11. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 217Combining the transfer function of the predictor ating region so that we obtain a similar closed-͑A8͒ and using the discrete convolution, the char- loop response for each one of them. Then, duringacteristic equation T ( z Ϫ1 ) can be written as fol- the operation, we vary J according to the operat-lows: ing region controlled at each sample. w ͑ 1Ϫz Ϫ1 ͒ ϩ ͚ q *˜ J z Ϫ jϪ1 j a 6.2. Performance analysis jϭ0 w N Finally, we analyze the effect of the filter over ϩ ͚ q* ͚ ˜ i z JϪiϪ j the overall closed-loop performance and compare j h jϭ0 iϭJϩ1 the predictive feedback controller with a standard MPC controller, such as the DMC. To carry out w N this analysis we compare the control action com- ϩ ͚ q * ͚ ͑ h i Ϫh i ͒ z ϪiϪ j j ˜ puted by both predictive controllers. jϭ0 iϭ1 The control actions generated by the predictive w ϱ feedback control law ͑27͒ are obtained by replac- ϩ͚ q* j ͚ h i z ϪiϪ j ϭ0. ͑36͒ ing the open-loop error e 0 ( J,kϪ j ) by their com- ˆ jϭ0 iϭNϩ1 ponents, assuming a step change in the setpoint,The stability of the closed-loop system depends on and combining with the predicted output ͑8͒, theboth the prediction time J and the filter param- result iseters. So, it may be tested by any usual stability w u ͑ k ͒ ϭu ͑ kϪ1 ͒ ϩ ͚ q * e ͑ kϪ j ͒criteria. Using the same procedure as in AppendixB, we can derive the following stability condition: j jϭ0 N ͚ iϭJϩ1 h N ͉˜ i ͉ ϩ ͚ ͉ h i Ϫh i ͉ ϩ iϭ1 ˜ ͚ͯ ͯϱ iϭNϩ1 h i Ͻa J . ˜ w ϩ ͚ q * ͓ P ͑ J,z ͒ ϪG p ͑ z ͔͒ u ͑ kϪ j ͒ . jϭ0 j ˜Note that this condition is the same as that derived ͑37͒for the single-prediction controller ͓Eq. ͑17͔͒.However, we should note that the parameters of In this equation we must observe that the last termthe filter q * , jϭ0,1,...,w affect the closed-loop j is a weighing contribution of the future open-loopstability. These facts look contradictory, because it deviations at time ( kϪ jϩJ ) T S ,is clear from the characteristic Eq. ͑35͒ that theclosed-loop stability depends simultaneously on ⌬y 0 ͑ J,kϪ j ͒ ϭy 0 ͑ J,kϪ j ͒ Ϫy ͑ kϪ j ͒ . ˆ ˆ ˜both. However, these results can be interpreted asfollows: the closed-loop stability for the predictive It only depends on the past control actions and thefeedback controller is obtained by the independent system model, so it states the effect of the pastselection of the prediction time J and the param- control actions on the future behavior of the sys-eters of the filter, such that they independently tem. This fact implies that it has significant influ-guarantee the closed-loop stability. These facts ence on the closed-loop performance when wemean that the prediction time J should be selected have to track the setpoint. However, this term haslike the single-prediction controller ͓Eqs. ͑17͒ to a negligible influence when we have to reject a͑19͔͒, and the filter must be tuned as there is no disturbance, because it has little information abouttime delay in the system, because it has been com- the disturbance. The two first terms of Eq. ͑37͒,pensated by the open-loop predictor. ordered from the left, are the time implementation Since the prediction time J can be fixed inde- of a reduced order controller of order w ͓6͔, andpendently of the filter’s parameters, we can vary it commands the system behavior during the distur-such that the closed-loop performance is im- bance rejection. It is clear now that the control lawproved. Varying J we modify the closed-loop set- ͑37͒ includes a direct feedback action based ontling time, speeding up or slowing down the sys- measured error.tem response. So, if we have to control a nonlinear Now, we compare the performance achieved bysystem we can choose a different J for each oper- the predictive feedback with that of an MPC con-
  12. 12. 218 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226Table 1Predictive controller parameters and results.Controller parameters f Set f Dist Nϭ60, Vϭ54, Uϭ4, ␭ϭ0.14 7.7164 3.0323 wϭ1; qϭ ͓ 0.7280 0.0987͔ 7.7061 2.5029 wϭ2; qϭ ͓ 0.7468 0.0825 0.0902͔ 7.6938 2.4595 wϭ3; qϭ ͓ 0.7614 0.0979 0.0792 0.0816͔ 7.7005 2.4229troller. Recalling Eq. ͑24͒ we have the control ac- point tracking and disturbance rejection. To evalu-tion generated by a standard MPC controller, ate the closed-loop responses we consider the fol- ͚ͭ ͮ V lowing linear plant: u ͑ k ͒ ϭu ͑ kϪ1 ͒ ϩ k j e͑ k ͒ e Ϫ50s jϭ1 Gp ͑ s ͒ ϭ , ͑ 150sϩ1 ͒͑ 25sϩ1 ͒ V Ϫ ͚ k j ͓ P ͑ j,z ͒ ϪG p ͑ z ͔͒ u ͑ kϪ1 ͒ . ˜ which was used by Rahul and Cooper to evaluate jϭ1 a tuning procedure for a DMC controller. The dis- ͑38͒ crete transfer function is obtained by assuming a zero-order hold at the input and a sampling timeIn this equation we can see that MPC controllers t S ϭ16. The tuning parameters for the DMC con-only use the last measured error and the two first troller are same as those using by Rahul and Coo-terms, ordered from the left, are a discrete PI con- per in his work [14] (see Table 1).troller. Like the predictive feedback controller, the The predictive feedback controller was designedlast term is a weighing contribution of the future by solving the basic tuning problem proposed byopen-loop deviations at time ( kϩ j ) t S , j Giovanini and Marchetti [19] with the objectiveϭ1,2,...,V. function Comparing Eqs. ͑37͒ and ͑38͒ we can see that Npredictive feedback controller uses more feedbackinformation than standard predictive controllers to f ϭ ͚ e 2 ͑ k ͒ ϩ␭⌬u 2 ͑ k ͒ , ͑39͒ kϭ1compute the control actions. Therefore the predic-tive feedback controller reduces the effect of dis- for given filter orders and a prediction time (Jϭ5).turbances more aggressively than any standard The order of the filter as chosen such that the re-MPC controller, and has better performance than sulting controllers include the predictive versionMPC, especially for disturbance rejection problem of popular PI and PID controllers (wϭ1,2,3). Theor important uncertainties present in the system. problem described to this point has a fast solution A final comment about the predictive feedback using an algorithm based on the descendent gra-can be made. From Eq. ͑37͒ it is easy to see that a dient method. The filter parameters obtained forpredictive feedback controller combines the ca- each controller are shown in Table 1.pacity of the predictive control algorithm for good The results obtained in the simulations are sum-setpoint tracking and time delay compensation, marized in Figs. 7 and 8. They show the perfor-with the classical use of the feedback information mance achieved by the predictive controllers nor-to improve the disturbance rejection. Furthermore, malized to DMC performance. It has beenfor controlling a nonlinear system the prediction computed astime J can be varied to improve the closed-loopperformance by modifying the closed-loop settling f DM C x Ϫ f PF xtime. ␳ x ϭ100 , xϭSet,Dist, f DM C x Example 6.1. In this example we compare theperformances achieved by a predictive feedback where f DM C x , f PF,x is the performance achievedcontroller and a standard MPC controller for set- by DMC and predictive feedback controllers for
  13. 13. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 219 Fig. 7. Performance comparison between DMC and predictive feedback, using index ͑39͒.setpoint tracking and disturbance rejection. Fig. 7 7. Simulations and resultsshows the normalized performance measured withEq. (39), while Fig. 8 shows the normalized per- Let us consider the problem of controlling aformance measured with the ISE index. The pre- continuously stirred tank reactor ͑CSTR͒ in whichdictive feedback controller exhibits a similar per- an irreversible exothermic reaction is carried outformance to the DMC controller (Fig. 7) for the at constant volume ͑Appendix C͒. This is a non-setpoint tracking. Although a better response is linear system previously used by Morningred etobtained by the predictive feedback controller al. ͓20͔ for testing predictive control algorithms.(Fig. 8), it uses more control energy than DMC Fig. 9 shows the dynamic responses to the follow-(Fig. 7). On the other hand, the predictive feed- ing sequence of changes in the manipulated vari-back controller shows an important improvement able q c : ϩ10, Ϫ10, Ϫ10, and ϩ10 l minϪ1 ,in the closed-loop performance for disturbance re- where the nonlinear nature of the system is appar-jection. A better response is again obtained by ent.predictive feedback controller (Fig. 8) and the Four continuous linear models are determinedcontroller uses a similar control energy than the using least-squared procedures to adjust the com-DMC (Fig. 7). Fig. 8. Performance comparison between DMC and predictive feedback using the ISE index.
  14. 14. 220 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 Fig. 9. Open-loop responses of the CSTR concentration to step changes in the coolant flow rate q C (t).position responses to the above four step changes actor using the polytope idea. These models arein the manipulated variable. Notice that those obtained by Z transforming the continuous trans-changes imply three different operating points cor- fer functions and assuming a zero-order-hold de-responding to the following stationary manipu- vice is included. This representation should be as-lated flow rates: 100, 110, and 90 l minϪ1 . Table 2 sociated to the M vertex models in the tuningshows the four process transfer functions obtained, problem formulation. In this application we stress the fact that the Ca ͑ s ͒ reactor operation becomes very sensitive once the ϭG P l ͑ s ͒ , lϭ1 – 4. q C͑ s ͒ manipulation exceeds 113 l minϪ1 . Hence, assum- ing that a hard constraint is physically used on theThey define the polytopic model associated to the coolant flow rate at 110 l minϪ1 , an additional re-nonlinear behavior in the operating region being striction for the more sensitive model ͑model 1 inconsidered. Table 2͒ must be considered for the deviation vari- Like in Morningred’s work, the sampling time able u ( k ) :period was fixed at 0.1 min, which gives aboutfour sampled-data points in the dominant time u 1 ͑ kϩi ͒ р10, 0рiрV. ͑40͒constant when the reactor is operating in the highconcentration region. Then, four discrete linear Besides, a zero-offset steady-state response is de-models are used for representing the nonlinear re- manded, then we include the following constraint:Table 2Vertices of the polytope model. Change Model obtained Model 1 Ϫ0.0008t 3 ϩ0.033 s 2 Ϫ0.018sϩ0.67 Ϫ0.5 s q C ϭ100, ⌬q C ϭ10 GP1͑s͒ϭ e s 4 ϩ1.92 s 3 ϩ30.35 s 2 ϩ21.49 sϩ153.7 Model 2 Ϫ1.3 10Ϫ5 s 3 ϩ0.0065 s 2 ϩ0.354 sϩ3.35 Ϫ0.5 s q C ϭ110, ⌬q C ϭϪ10 GP2͑s͒ϭ e s 4 ϩ10.5 s 3 ϩ101.37 s 2 ϩ334.89 sϩ834.6 Model 3 6.7 10Ϫ6 s 3 Ϫ0.0055 s 2 ϩ0.652 sϩ9.35 q C ϭ100, ⌬q C ϭϪ10 GP3͑s͒ϭ e Ϫ0.5 s s 4 ϩ28.45 s 3 ϩ324.67 s 2 ϩ1737.15 sϩ3718.6 Model 4 Ϫ1.07 10Ϫ4 s 3 ϩ0.0256 s 2 ϩ0.143 sϩ0.457 Ϫ0.5 s q C ϭ90, ⌬q C ϭ10 GP4͑s͒ϭ e s 4 ϩ9.58 s 3 ϩ49.69 s 2 ϩ128.05 sϩ178.4
  15. 15. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 221Table 3 time J. It is fixed such that the robust stability ofNominal CSTR parameter values. the system is guaranteed. The prediction time sat- Parameter Nomenclature Value isfiesMeasured Ca 0.1 mol lϪ1 J stb ϭmax͑ J 1 ;J 2 ;J 3 ;J 4 ͒ ϭ10. ͑42͒ concentrationReactor T 438.5 K Notice in this case that it is the polytope that must temperature be shaped along the time being considered. HenceCoolant flow rate qC 103.41 l minϪ1 the objective function necessary for driving theProcess flow rate q 100 l minϪ1 adjustment must consider all the linear models si-Feed concentration Ca O 1 mol lϪ1 multaneously. At a given time instant and operat-Feed temperature TO 350 KInlet coolant T CO 350 K ing point, there is no clear information about temperature which model is the convenient one for represent-CSTR volume V 100 l ing the process. This is because it depends notHeat-tranfer term hA 7.0ϫ105 only on the operating point but also on which di- cal minϪ1 KϪ1 rection the manipulated variable is going to move.Reaction rate k0 7.2ϫ1010 minϪ1 The simpler way to solve this by proposing constant M VActivation energy E/R 1.0ϫ104 KϪ1Heat of reaction ⌬H Ϫ2.0ϫ105 fϭ͚ ͚ ␥ l ͓ e l ͑ k ͒ 2 ϩ ␳ ⌬u l ͑ k ͒ 2 ͔ , ͑43͒ lϭ1 kϭ1 cal molϪ1Liquid densities ␳,␳c 1.0ϫ103 g lϪ1 where the time span is defined by Vϭ200. TheSpecific heats C p ,C pc 1.0 cal g KϪ1 control weight ␳ was fixed in a value such that the control energy has a similar effect as errors in the tuning process ( ␳ ϭ0.01) . Since in this applica- y͑k͒ р1.05 r ͑ k ͒ , 0рkрN, tion we found no reason to differentiate the mod- els, we adopt ␥ l ϭ1, l෈ ͓ 1,M ͔ . The problem de- ͉ e͑ k ͉͒ р5.0 10Ϫ4 , 50рkрN. ͑41͒ scribed to this point has a rapid numerical solution using an algorithm based on the gradient method.This assumes that the nominal absolute value for The parameters obtained are the following:the manipulation is around 100 l minϪ1 and thatthe operation is kept inside the polytope whose q 0 ϭ0.2313; q 1 ϭϪ0.1550;vertices are defined by the linear models. Now, we define the parameters of the predictive q 2 ϭϪ0.2531; q 3 ϭ0.2157. ͑44͒feedback controller to tune the filter parametersusing the method proposed by Giovanini and Mar- Morningred et al. ͓20͔ have previously workedchetti ͓19͔. The filter size w is arbitrarily adopted with this reactor model for testing different alter-( wϭ3 ) and the predictor of the controller P ( J,z ) natives of predictive controllers and confrontedis built using the nonlinear model ͑C1͒, assuming the results with the responses obtained using a PIthe nominal parameters value ͑Table 3͒. An adap- controller whose parameters were adjusted by thetive Runge-Kutta integration scheme is used to ITAE criterion; thus we used the same settings: thecompute the system output prediction, with the gain value 52 l2 molϪ1 minϪ1 and the integrationinitial state given by X ( k ) ϭ ͓ Ca ( k ) T ( kϪ5 ) ͔ at time constant 0.46 min. The simulation tests areeach sample. Since the inlet coolant temperature also similar to Morningred’s work and consists ofT CO ( t ) is measurable, we include it in the predic- a sequence of step changes in the reference valuetor to improve the rejection to this disturbance. and a sequence of load changes in the feed stream Because the controller predictor is built using concentration and the refrigerant inlet temperature.the nonlinear model, it is reasonable to assume Fig. 10 shows the results obtained when com-that there are no uncertainties. Therefore we build paring the discrete controller with the mentionedthe open-loop predictor of the tuning problem ͓19͔ PI. The setpoint was changed in intervals of 10using the same model as that used to simulated the min. from 0.09 to 0.125 mol lϪ1 , returns to 0.09,system output. Finally, we choose the prediction then steps to 0.055 and returns to 0.09 mol lϪ1 .
  16. 16. 222 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226Fig. 10. Closed-loop responses of the CSTR concentration to a sequence of step changes in the setpoint using the predictivecontroller with one-side constraint in the manipulation and a PI controller adjusted with ITAE.The superior performance of the discrete control- bility is observed in the discrete controller sug-ler is obtained through a vigorous initial move- gested in this paper.ment in the manipulated variable, which, however, Fig. 13 shows the manipulated movements cor-does not overcome the 110 l minϪ1 limit as shown responding to the repsonses in Fig. 12. In this fig-in Fig. 11, but shows more movements than the PI. ure we see that the excursion of q C ( t ) is more Fig. 12 shows the results obtained when com- important in the case of the PI, but smoother.paring the discrete controller with the mentionedPI under load changes. For testing the disturbance 8. Conclusionsrejection the following sequence of changes aremade: first the feed stream concentration Ca O ( t ) A new method to design predictive controllerschanges from 1 to 1.05 mol lϪ1 and 10 min later for linear SISO systems has been presented in thisthe refrigerant temperature T CO ( t ) goes down work. It uses only one prediction of the system10 °C; then Ca O ( t ) and T CO ( t ) return to the output J time intervals ahead to compute the cor-original values, with a 10-min difference between respondent future error. Then, the predictive feed-them. A better disturbance rejection capa- back controller is defined by introducing a filter Fig. 11. Manipulated movements corresponding to the responses in Fig. 10.
  17. 17. Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226 223Fig. 12. Closed-loop responses of the CSTR concentration to a sequence of load changes using the predictive controllers.which weights the last predicted errors. In this analyzed. An extensive analysis of closed-loopway, the resulting control action is computed by performance, compared with standard MPC con-observing the system future behavior and the trollers, have been also carried out.present and past errors. These features enable the In spite of the results obtained in his work, sev-predictive feedback controller to combine the ca- eral questions about extension to multivariablepacity of predictive control algorithm for good set- systems and how we can address on-line con-point tracking and time delay compensation, with straints in the input and output variables still re-the classical use of the feedback information to main open as future research topics. A future workimprove the disturbance rejection. can also include an on-line tuning of the predictive The character of these controllers is governed by feedback parameters such that the performance re-one parameter, the prediction time, which directly mains optimal and the constraints would be full-influences the speed of closed-loop response. filled for every sample.Some simple criteria for its selection are provided:they guarantee the robust stability of the closed-loop system. The predictive feedback controller Appendix A: Predictor transfer functionhas additional tuning parameters: the parametersof the filter. Robust stability and closed-loop per- Eq. ͑5͒ defines a realizable J -step ahead predic-formance issues of these controllers have been tor in the discrete time domain. The predictor is Fig. 13. Manipulated movements corresponding to the responses in Fig. 12.
  18. 18. 224 Leonardo L. Giovanini / ISA Transactions 42 (2003) 207–226realizable since only past inputs to the system areused to computed the future behavior. Expandingthis equation results in ˆ h ͫ h J y ͑ J,z ͒ ϭy ͑ z ͒ ϩ ˜ 1 z Ϫ1 ϩ ͚ ˜ i z Ϫ1 ˆ0 iϭ2 N y 0 ͑ J,k ͒ ϭy ͑ 0,k ͒ ϩ ͚ ˜ i ⌬u ͑ kϩ1Ϫi ͒ ˆ ˆ h iϭ2 h ͬ ϩH ͑ J,z ͒ Ϫ ͑ ˜ 1 z Ϫ1 ϩH ͑ 1,z ͒ z Ϫ1 ͒ u ͑ z ͒ , ͑A6͒ N ϩ¯ϩ ͚ iϭJϩ1 ˜ i ⌬u ͑ kϩJϪi ͒ , h where we can identify the Jth coefficient of step response,and taking the Z transform gives J ˜ Jz a Ϫ1 ϭh 1 z Ϫ1 ϩ ͚ ˜ i z Ϫ1 , ˜ h y 0 ͑ J,z ͒ ϭy ͑ z ͒ ˆ ˆ iϭ2 ϩ ͚ͫ N iϭ2 ˜ i z 1Ϫi ϩ¯ϩ h N ͚ ˜ i z JϪi iϭJϩ1 h ͬ and the plant model H ͑ 0,z ͒ ϭh 1 z Ϫ1 ϩH ͑ 1,z ͒ z Ϫ1 ϭG p ͑ z ͒ . ˜ ˜ Ϫ1 ϫ ͑ 1Ϫz ͒u͑ z ͒. ͑A1͒ Hence the expression ͑A6͒ can be writtenDefining the following function: ͭ y 0 ͑ J,z ͒ ϭy ͑ z ͒ ϩ ͓ ˜ J z Ϫ1 ϩH ͑ J,z ͒ ϪG p ͑ z ͔͒ u ͑ z ͒ , ˆ ˆ a ˜ N ͑A7͒ H ͑ J,z ͒ ϭ ͚ iϭJϩ1 ˜ i z JϪi , h 0рJрNϪ1 and, since 0, JϭN, ͑A2͒ y ͑ z ͒ ϭG p ͑ z ͒ u ͑ z ͒ , ˆ ˜Eq. ͑A1͒ can be written as the Z transform of the J-step ahead single predic- tor is given by y 0 ͑ J,z ͒ ϭy ͑ z ͒ ϩ ͓ H ͑ 1,z ͒ ϩ¯ϩH ͑ J,z ͔͒ ˆ ˆ ϫ ͑ 1Ϫz Ϫ1 ͒ u ͑ z ͒ . ͑A3͒ P ͑ J,z ͒ ϭa J z Ϫ1 ϩH ͑ J,z ͒ , ˜ ͑A8͒Note that there is a recursive relationship, i.e., y 0 ͑ J,z ͒ ϭ P ͑ J,z ͒ u ͑ z ͒ . ˆ ͑A9͒ H ͑ m,z ͒ ϭh mϩ1 z Ϫ1 ϩH ͑ mϩ1,z ͒ z Ϫ1 . ͑A4͒ ˜Then, combining Eqs. ͑A3͒ and ͑A4͒ and rear- Appendix B: Robust stability condition forranging gives single-predictive controller y ͑ J,z ͒ ϭy ͑ z ͒ ϩ ˆ0 ˆ ͚ͫN iϭ2 ˜ i z Ϫ1 ϩH ͑ J,z ͒ h The characteristic closed-loop equation T ( z Ϫ1 ) for the single-prediction controller is given by ϪH ͑ 1,z ͒ z Ϫ1 ͬ u͑ z ͒. ͑A5͒ T ͑ z Ϫ1 ͒ ϭa J ϩH ͑ J,z Ϫ1 ͒ ϩGp ͑ z Ϫ1 ͒ ϪG p ͑ z Ϫ1 ͒ . ˜ ˜ ͑B1͒ Using the discrete convolution and Eq. ͑A2͒, theAdding and substracting ˜ 1 z Ϫ1 and operating h last expression can be written in the followinggives way: