A Development of Tools For Monitorization and Control of Multivariable Neurocontrolled Systems with Application to Distillation Columns

Transcription

1 A Deveopment of Toos For Monitorization and Contro of Mutivariabe Neurocontroed Systems wit Appication to Distiation Coumns J Fernandez de Canete, S Gonzaez-Perez, P de Saz Orozco Dpt of System Engineering and Automation ETSI Industriaes, Paza E Ejido s/n 293 Maaga, SPAIN E-mai: canete@isaumaes, spg@isaumaes, itfs@yaoocom Abstract Te issues raised in advanced process contro appications ead contro engineers to be aware of te need of integrating advanced information processing capabiities in te process contro oop In tis paper a SIMUINK based too as been deveoped for anaysis and design of mutivariabe neura based contro systems Tis too as been appied to te contro of a ig purity distiation coumn aso modeed in SIMUINK incuding non inear ydrodynamic effects Aso abiew environment as been empoyed as a grapica user interface for monitoring te neurocontroed distiation coumn, by visuaizing bot te cosed oop performance and te user seected contro conditions Te proposed contro sceme offers an optima response for bot teoretica and practica caenges posed in process contro tas, in particuar wen bot, te quaity improvement of distiation products and te operation efficiency in economica terms are considered Introduction Artificia neura networs ave been used extensivey for a number of process engineering appications suc as sensor data anaysis, faut detection and noninear pant identification and contro, specificay of distiation coumn and reactors in cemica engineering Focusing on te contro probem, severa contro scemes based on nowedge of te pant mode ave been reported, suc as generaized predictive contro [], inverse mode contro [2] and adaptive contro [3] among oters Te ac of toos for te design of controers based on neura networ modes is particuary pronounced Te reason for tis migt be tat deveopment of generic software for contro system design is reativey difficut as severa types of contro designs exist A few non commercia toos for system identification and contro system design ave become avaiabe One of tese is te system identification and contro MATAB-SIMUINK too [4] wic offers an usefu guided user interface (GUI) for predictive contro, mode reference adaptive contro and feedbac inearization contro, a based on neura networs In addition, [5] ave deveoped two tooset for use wit MATAB, termed NNSYSID and NNCTR for neura networ based identification and contro of noninear systems Tese tooset aow te user to coose among severa designs, suc as direct inverse contro, interna mode contro, feedbac inearization and predictive contro among oters However, tese tooset are appied ony to SISO noninear systems, being terefore invaid to be extended to genera MIMO contro probems as is frequenty usua in process engineering On te oter and abiew [6] and SIMUINK are propery two widey used grapica code deveopment environment wic aow system eve deveopers to perform rapid prototyping and testing Unie text-based programming anguages, suc as C, MATAB and Java, grapica programming invoves boc-based code deveopment and offers a more intuitive approac to designing contro systems Besides, bot environments can be ined togeter troug te abiew simuation interface tooit (SIT) aowing contro engineers to custom user interface to interactivey verify SIMUINK modes and easiy depoy tese modes to rea-time ardware

2 for contro prototyping and ardware-in-te-oop testing, wit proven resuts in industria appications [7] For many reasons, distiation remains te most important separation tecnique in cemica process industries around te word [8] Terefore, improved distiation contro can ave a significant impact on reducing energy consumption, improving product quaity and protecting environmenta resources However, distiation contro is difficut because it is usuay noninear, non-stationary, interactive, and is subject to constraints and disturbances In tis paper a MATAB-SIMUINK based too as been deveoped for anaysis and design of mutivariabe neura based contro systems Tis too as been appied to te contro of a ig purity distiation coumn modeed aso in SIMUINK incuding non inear ydrodynamic effects Aso, abiew environment as been empoyed as a grapica user interface for monitoring te neurocontroed distiation coumn, by visuaizing bot te cosed oop performance and te user seected contro conditions, were te front pane as been designed mimicing te distiation coumn contro sceme SIMUINK neura er trim 2do trim 3er trim 4to trim Resuts SIMUINK neura contro too abiew front pane Piot-scae distiation coumn SIMUINK mode Fig Sceme for monitoring and contro of te distiation coumn

3 In section II we describe te overa structure of neurocontroed system wie te distiation coumn mode programmed in SIMUINK is summarized in section III In section I we present te too for anaysis and design of mutivariabe neura based contro systems Te design of te monitoring system under abiew is described in section, and computer simuation resuts for a 9 pates neurocontroed coumn are aso sown in section I Finay, concusions and future wors are aso discussed in section II 2 Description of te Contro System te remainder is taen as te distiate product D Part of te iquid fow out of te bottom of te coumn is vaporized in a reboier and sent bac to te bottom of te coumn, wie te remainder is taen as te bottom product B Te coumn consists of a N bubbe cap trays Te overead vapour is totay condensed in a water cooed condenser (tray N+) wic is open at atmosperic pressure Te reboier (tray ) is eated eectricay, and te preeated feed stream enters te coumn at te feed tray as saturated iquid Te process inputs tat are avaiabe for contro purposes are te eat input to te reboier Q and te refux fowrate N+ (Fig 2) Te proposed arcitecture for neura contro of te distiation coumn is sown (Fig ) Te system is composed by a) A SIMUINK mode of te distiation coumn, based on mass and entapy baances; b) An interactive too running under SIMUINK for identification and contro based on neura networs, using I/O data form te distiation pant and c) A grapica user interface for monitoring te neurocontroed distiation coumn based on abiew Te data taen from te SIMUINK distiation coumn mode wi be used to train bot te neura networ mode and controer, wic in turns once trained wi feed te distiation pant to operate under woring specifications Te outcome of bot modues wi be dispayed at front pane of abiew to sow te user te state of te pant at any time Te overa system aows aso te controing of a experimenta system instead of te simuation mode, using instead te measures form te sensors of te coumn in order to train bot te neura mode and controer under abiew interface 3 Distiation Coumn Mode Te distiation coumn is used for te separation of a binary mixture of metano and n-propano wic enters as a feed stream wit fow rate F, composition F and entapy q between two sections (a rectifying section and a stripping section) Mass transfer occurs between te vapour fowing up and te iquid fowing down te coumn Te vapour exiting at te top of te coumn is condensed, and part of te resuting iquid fow is returned at te coumn at te top (refux), wie Fig 2 Sceme of te binary distiation coumn In tis section we wi describe te mode of te distiation coumn deveoped by [9], composed by te mass, component mass and entapy baance equations used as basis to impement te SIMUINK diagram 3 Mass Baance Te differentia equations tat determine te cange of te moar odup n of te trays are given by n = + + Ony vaid for =,,, +,, N trays, were is te moar vapour fux eaving te - t tray and is te iquid fux eaving te -t tray For te feed tray te equation is given by

4 n + + F = + and te corresponding for te condenser and reboier are n = N N + N + n = 32 Component Mass Baance D B For te most voatie component, te composition baance wi foow te equation ( n ) = Y Y vaid for =,,, +,, N were is te iquid concentration of te most voatie component and Y is te composition of te vapour fow out of te -t tray For te feed tray we wi ave ( n + ) = + Y Y + F wie for te reboier and condenser we ave ) ( n = Y + B ( n N + N + ) = NYN N + N + D N + 33 Entapy Baance We wi assume tat te condenser pressure is fixed to te outside atmosperic pressure, and tat te pressure in te trays can be cacuated under te assumption of constant pressure drop from tray to tray P = P + + P, and te tray temperatures are impicity defined by te assumption tat te tota pressure is equa to te sum of te partia pressures, s s P = P ( T ) + ( ) P2 ( T ), wit T ods for te temperature of te -t tray, and were te partia S pressures P can be evauated as F + s P = ( T ) exp A B T + C wit different vaues of te constants according to te iquid we are evauating Te non-ideaity of te trays and oter unmodeed effects are accounted in te tray efficiencies α Wit tese tray efficiencies we can cacuate te composition of te vapour fow out of te -t tray as a inear combination of te idea vapour composition on te tray and te incoming vapour composition of te tray beow by Y starting wit Y P ( T ) s = α P + ( α ) Y P ( T ) s = P We wi formuate ten te entapy baances in order to determine te vapour streams tat eave eac tray as foows n + n T = T + vaid for =,,, +,, N, were is te entapy wit indexing to te pase ( refers to iquid and to vapour) n and For te feed tray and te condenser we wi ave + n + + T + + n Q Q oss + n T = + T T = + B + F F Te entapies and are given by

5 = ( T ) + ( ) 2( T ) 4 Neura Modeing and Contro Too were = Y ( T, P) + ( Y) 2 ( T, P) can be evauated as = C 2 ( ( T T ) + ( T T ) + ( T ) ) 3, 2, 3, T wit different vaues for te constants according te iquid were referring to, and te vapour entapies can be evauated as 3 c P T = + RT c c P T 7 T T + + Ω T a b c d e c c T T T c T + f T c 7 wit coefficients according to te iquid or vapour we are considering in eac case Neura networs ave become a popuar too for identification and contro of unnown noninear systems Te Neura Networ Toobox commerciaised wit MATAB [4] is intended to serve as a generapurpose pacage for tis tas, but te efficient expoitation of tese services drasticay depends on te programming sis and experience of te users In tis context, te design of MIMO Contro Too running under SIMUINK and speciay conceived for buiding neura-networ-type modes becomes a more attractive perspective for deveoping appications tan writing aborious MATAB codes in a cassica manner In te context of tis paper, we present a design too for neura based modeing and contro of noninear systems, buit aso to anayze te stabiity contro probems associated (Fig 4) We ave appied tis metodoogy for a ig purity distiation pant, wic is ceary a MIMO noninear system Te SIMUINK mode corresponding to tese equations is depicted in Fig 3 sowing te top eve modeing description Fig 3 SIMUINK mode of te distiation coumn Fig 4 Sceme of te neura networ design contro too

6 4 Neura Identification Too of distiate composition for refux and eating fows signas seected in te operating ranges Te data for training bot te neura networ mode and controer were obtained from dynamic simuations using te SIMUINK mode aready deveoped in section III Te refux rate o and eat fow Q were used as pant inputs togeter wit top and bottom compositions D and B, wie feed variabes (F, F, q) ave been treated as process disturbances Te training set for identification comprised 2 data points beonging to te open oop operating range for pant inputs refux fowrate o (-5E-6 m 3 /) and eat fow Q (-2 J/s) for fixed feed rate conditions F = E-6 m 3 /, F = 3, and q = An additiona data set consisting of 5 data points was used to test te neura networ mode afterwards For training pattern generation we assume an initia steady state for te coumn after a start-up process Te identification tas was made using te evenberg-marquardt agoritm for a neura networ wit two ayers in a tansig - purein activation function sceme, by using automatic target vector generation wit te pant outputs Wit tese resuts we obtained an optimum 2--2 networ SIMUINK boc tat can be used for running severa experiences (Fig 5) Fig 5 SIMUINK structure for training te neura networ mode Figure 6 sows te comparison between te actua (SIMUINK) and te neura networ predicted vaues Fig 6 Prediction for distiate composition of te neura mode for o = 3E-6 m 3 /, Q = 2 J/s 42 Neura ControToo Tis too as been impemented using two approaces, direct design and indirect design [5] In te direct design metod te controer is in itsef a neura networ, wie in te indirect design te controer is not a neura networ at a, but te design is based on a neura networ mode of te pant to be controed Besides, te system can be specified eiter as SIMUINK mode, a MATAB fie containing te agebraic differentia equations or a neura networ mode of te system Te training set for contro comprised 5 data points beonging to te cosed oop operating range for desired and actua top and bottom compositions vaues D (-) and B (-) An additiona data set consisting of 2 data points was used to test te neura networ controer aso Te contro tas was made using te evenberg- Marquardt agoritm for a neura networ wit two ayers in a tansig - purein activation function sceme, couped wit a neura identification networ previousy trained wit te Neura Identification Too, so as to adjust te networ controer parameters using te propagation of te output pant error troug te neura networ mode [] We obtained an optimum networ SIMUINK boc for te neura controer, wit top and bottom composition errors as inputs and refux rate and eat fow as outputs (Fig 7)

7 Fig 7 SIMUINK structure for training te neura networ controer 5 Monitoring of te Distiation Coumn One of te goas of te present wor is to deveop a system tat aows te monitoring in ine and manipuate te variabes tat tae part in te distiation process, as we as to interact wit te neurona too we ave deveoped in a SIMUINK environment, Terefore it wi be of great interest to deveop aternatives to sove tis probem in an economic, efficient and reiabe way Te soution we propose is a grapic user interface deveoped in Nationa Instrument s grapica programming anguage abiew [6] In tis way, te observation and manipuation of te variabes is carried out using virtua instruments wie te connection between te virtua instruments and te neura too is soved using te Simuation User Interface ibrary, wic aows to deveop, prototype, and test contro systems using modes deveoped in te SIMUINK simuation environment (Fig 8) Fig 8 Interaction between abiew and SIMUINK In tis way, te programming anguage abiew was used for te tass of acquisition, anaysis and presentation of data in a friendy way to te user and aso provides a efficient too for te communication wit te neura too So te soution we propose ead us to a fexibe and versatie too, being an efficient, economic and simpe way to monitorize a dynamic pant and bot to mode and contro a distiation coumn by using te neura too (Fig 9)

8 6 Resuts In order to sow te vaidity of te proposed contro sceme, we ave seected a binary ig-purity distiation coumn, for separating a mixture of metano and n-propano, wic consist in 4 bubbe caps trays wit eated eectricay reboier and water refrigerated tubuar condenser In Fig, we sow ow te neura controer exibit adequate contro action to compensate a dua step cange in distiate composition from 99% to 98% and bottom composition from 985% to 975% in t = 4 s, togeter wit a cange in feed composition from 5% to 8% Canges in te refux and eat fows are determined by te neura networ mode-based controer for te coumn Fig 9 Frontpane of te monitoring too Fig Response of te distiation contro system to step canges in top and bottom compositions and cange in feed composition

9 7 Concusions and Future Wors A SIMUINK based too as been deveoped for anaysis and design of mutivariabe neura based contro systems Tis too as been appied to te contro of a ig purity distiation coumn aso modeed in SIMUINK incuding non inear ydrodynamic effects Aso abiew environment as been empoyed as a grapica user interface for monitoring te neurocontroed distiation coumn Te resuts obtained demonstrate te potentia use of tis contro strategy in tis fied Figure ab-scae piot distiation coumn DETAAB DC-SP Future wors are directed towards te appication of te described tooset to a experimenta distiation coumn as is actuay being made by te System Engineering and Automation Group as part of te researcing project DPI (fig ), wic utiizes te same sceme as Fig but repacing te simuation mode by an experimenta distiation coumn DETAAB DC-SP, wose tecnica caracteristics are dispayed at web page beow (in Spanis)ttp://wwwisaumaes/C4/Contro%2Neurob orroso/document%2ibrary/indextm) At te same time, te stabiity issued invoved in te neura contro tas are aso object of researc, since noninear and mutivariabe dynamics are present 8 References [] McMurray, J, Himmebau, D, Identification of a paced distiation coumn for contro via artificia neura networs, Proc ACC, San Francisco, CA, pp , 993 [2] iong, Q, Jutan, A, Grey-box modeing and contro of cemica processes, Cem Eng Science, o 57, pp 27-39, 22 [3] Hussain, M A, Review of te appications of neura networs in cemica process contro Simuation and on-ine impementations, Artificia Inteigence in Eng, o 3, pp 55-68, 999 [4] Demut, H, Beae, M, Hagan, M Neura Networ Toobox for use wit MATAB Te Matwors, 26 [5] Norgaard, M, Ravn, O, Pousen, N, NNSYSID and NNCTR toos for system identification and contro wit neura networs, Computing and Contro Engineering Journa, o 23, pp29-36, 2 [6] Bisop, R earning wit abiew 7 Express, New Jersey, Prentice Ha, 24 [7] Muamad, NA, Ai, S abiew wit fuzzy ogic controer simuation pane for condition monitoring of oi and dry type transformer, IEEE Trans on Engineering, Computing and Tecnoogy, o 4, pp 87-93, 26

10 [8] uyben, P Process modeing, simuation and contro for cemica engineers McGraw Hi, Cemica Eng Series, 99 [9] Die, M, Usu, I, Findeisen, R, Rea-time optimization for arge scae processes: Noninear predictive contro of a ig purity distiation coumn, On ine Optimization of arge Scae System: State of te Art, Springer-erag, 2 [] Norgaard, M, O Ravn, NK Pousen and K Hansen Neura Networs for Modeing and Contro of Dynamic Systems Springer erag, 2