Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890 Teperature Control Strategy of the Heatng Furnace Used n Adsorpton Desulfurzaton Devce Haozh Lu a, Hayun Chen, Daoka Yang School of Mechatronc Engneerng, Southwest Petroleu Unversty, Chengdu 610500, Chna Abstract a 7607164@qq.co Analyzed the cascade PID control schee used to control the teperature of the adsorpton desulfurzaton devce (S Zorb) heatng furnace and ts bg lag features, ths paper propose a predctve control odel that cobne cascade PID and RBF neural network predctve odel to control the teperature of the S Zorb heatng furnace. Ths paper takes RBF neural network as predctve odel to control the outlet teperature of the heatng furnace, neural network predctve value as the feedback sgnal of the an loop of the cascade syste and uses off-lne learnng to buld the neural network predctve odel,on-lne update to odfy the paraeters of the RBF neural network tely. The sulaton results showed that the schee could effectvely control soe dynac perforance, teperature overshoot, adjustent te and steady error, and control effects were better than those n cascade PID control schee. The cascade-rbf neural network syste has better control character and stronger robustness to overcoe the dsturbance and odel satch on S zorb heatng furnace. Keywords S Zorb heatng furnace; Cascade - RBF; Predctve control; The teperature control. 1. Introducton Wth the developent of autooble ndustry, the ncrease of the consupton of fossl energy, the envronent proble becoes ore and ore serous. Producton of clean gasolne wth low sulfur becoes the focus of the work of petrochecal enterprses. Ths paper takes the S zorb devce used n SINOPEC to product the low sulfur gasolne as the research object. S zorb was developed prarly for adsorpton desulfurzaton technology of FCC gasolne fracton by ConocoPhllps. Ths technology has advantages of hgher desulfurzaton rate, lower octane nuber loss and lower operaton cost. The heatng furnace used n S zorb s used to heat hydrogen-gasolne xture and crculatory hydrogen, and ts outlet teperature fluctuaton wll deterne the reactor s desulfurzaton qualty and octane nuber loss whch are portant techncal ndcators. So, the research on furnace teperature control strategy used n S zorb devce has portant sgnfcance and necessty. At present te, the strategy of teperature control of the furnace ostly adopt schee of controllng the outlet teperature of the furnace. Also, S zorb adopt the pressure-teperature cascade control schee to control the teperature of the furnace, take outlet teperature control as the an loop of the schee, take the pressure of fuel gas control as the secondary loop. Even though the cascade control strategy has preferable control perforance to suppress dsturbance n secondary loop, proves the dynac perforance of the controlled syste and frequency of the syste, but can t overcoe ts bg lag. So the teperature control has large constant te and bg lag. Ths paper propose a predctve control odel that cobned cascade PID and RBF neural network predctve odel to control the teperature of the S-Zorb heatng furnace, took RBF neural network as predctve odel to control the outlet teperature of the heatng furnace and neural network predctve value as the feedback sgnal of the an loop of the cascade syste. Fnally, a coparson between cascade PID and cascade-rbf neural network syste wth sulaton results s sulated by MATLAB. 1
Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890. The research on teperature control used n S zorb furnace based on cascade RBF neural network predctve control Secton Headngs.1 Introducton of the S zorb heatng furnace control S zorb heatng furnace s convecton-radaton vertcal cylndrcal type heater shown n fg.1. frstly, ateral pup send Hydrogen-gasolne xture to heat transfer Hydrogen-gasolne xture s conveyed fro heat exchanger to convecton chaber after preheated by heat exchanger, then been sent to bottle of the furnace after heated by radant col, fnally the xture was sent to reactor for adsorpton desulfurzaton reacton. The teperature of ateral has been heated copletely s portant paraeter of S zorb technology and ust be control precsely to ensure reacton effectvely n reactor. So, S zorb devce choose outlet teperature of Hydrogen-gasolne xture as an control paraeter. Fg.1 S-Zorb heatng furnace teperature control schee Man Secondary Control Valve f1 Fuel Gas Presser Furnace f Materal Teperature Presser Sensor Teperature Sensor Fg. Outlet teperature of the ateral - the fuel gas pressure cascade control block dagra As shown n Fg., the control schee of S zorb heatng furnace adopts outlet teperature of ateral - the fuel gas pressure cascade control strategy at present te. The schee takes teperature of hydrogen-gasolne xture as an control paraeter, the pressure of fuel as senor loop control paraeter, PID as controller. Interference produced by fuel gas ppe depressurzaton when lock hopper convey catalyst, Interference produced by ntal teperature of the hydrogen-gasolne xture, both of the wll be elnated by controllers of an loop and senor loop. The pressure of fuel s easured by ntellgent pressure sensors; teperature of xture s easured by k theral couples. The schee has large constant te and bg lag feature.. Recursve predctve control odel based on neural network It s dffcult to obtan the accurate odel on outlet teperature control of S zorb heatng furnace. Ths paper proposes to use neural network to buld predctve odel and take neural network advantage of
Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890 strong approxaton ablty of nonlnear functon. Neural network can adapt to nonlnear and te-varyng characterstcs of heatng furnace, avod nfluence of odel satch on control perforance. As shown n Fg.3, ths paper buld the predctve odel by takng use of neural network, takes neural network predctve value as the feedback sgnal of the an loop of the cascade syste. Fg.3 s recursve predctve control odel based on neural network. where represents the set pont of syste, s output of syste, U(k)s regulatng varable, s predctve value of k step. Rk ( ) Uk ( ) Controlled Yk ( ) Plant Yk ( ) yk ( ) Rk ( ) Neural Network yk ˆ( ) Predcton - + Fg.3 Neural network predctve control odel The SISO nonlnear syste could be present by followng dscrete odel. respectvely output and nput value of syste. 3 yk and u k 1,,, 1, represent y k f y k y k n u k d u k d (1) Predctve odel based on neural network could be present as followng, functon. Predctve value of the k 1,,, 1, f NN s neural network yk ˆ( ) fnn y k y k n u k d u k d () d step are 1,,, 1, yˆ ( k b) fnn y k b y k b n u k b d u k b d (3) (3) Usng the predctve value to replace nput value could obtan followng equaton. 1,,, yˆ ( k b) f ˆ ˆ NN y k b y k b n u k b d 1, u k b d ( 4) (4) It s concluded that recursve predctve control odel based on neural network can get any step output of syste just gvng the odel of syste. Due to the recursve way need prevous forecasts as nput, the fnal predctve value wll accuulate error wth ncreasng steps when any predctve value s not accurate. Therefore, n order to eet the requreents of controllng syste real-te, t s necessary to use on-lne update to odfy the paraeters of the neural network every step..3 Cascade - RBF neural network predctve control schee Radal Bass Functon (RBF) neural network s a three-layer network put forward by Meoody, ts nput layer node nuber s equal to the nuber of ndependent varables n the research proble, ts hdden layer take radal bass functon as transfer functon, ts output layer s a lnear adder. So the output of the network s a weghted lnear addton of hdden layer outputs, the network weght can be obtaned by way of solvng lnear equatons. Copared wth BP neural network, Radal Bass Functon (RBF) neural network has stronger capablty of nonlnear approxaton wth hgh reconstructng accuracy and faster tranng rate. Cascade - RBF neural network predctve control schee s proved fro cascade PID control schee of S zorb heatng furnace. Copared wth the forer, Cascade - RBF neural network predctve control schee ntroduce RBF neural network predctve controller to the forer and take the predctve value as feedback paraeters. Fg.4 s control block dagra. In Fg.4, the schee
Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890 takes ncreent PID control arthetc as controller; s the error between settng value and predctve value yp k b ;, are nterference added to senor loop and an loop respectvely; rk s the error between predctve value and easured value. f 1 f k ek ( ) yr k yr () k y ( ) p kb f 1 ek ( ) Man Secondary Control Valve Pressure Sensor RBF Neural Network Fuel Gas Pressure Furnace f Materal Teperature rk () Fg.4 Cascade - RBF neural network predctve control schee.3.1 The establshent of off-lne odel The establshent of RBF neural network off-lne odel s ostly obtanng paraeters of RBF neural network. The RBF network confguraton s forulated as a nzaton proble wth respect to the nuber of hdden layer nodes, the center locatons and the connecton weghts. the relatonshp between nput and output could be present by follows. yx ( ) wg( X c, ) (5) X c The bass functons use Gaussan functon: exp( ). Where 4 c s center poston of bass functon, s wdth value of basc functon. Obtanng the center locatons of hdden layer usually take use of clusterng ethod, such as k-eans clusterng algorth. The wdth value of basc functon can be obtaned by follow equaton: d (6) K Where d s the axu dstance between cluster centers, K s the nuber of clusterng center. The coplex-valued weghts between hdden and output layer are updated by solvng lnear syste based on fndng the coplex-valued weghts between nput and hdden layer Where W s weght, G s pseudo nverse of the atrx G. Matrx G s deterned by correspondng output of hdden layers W G y (7) G gj x j c (8) gj exp( ) The establshent of RBF neural offlne predcton odel s copleted..3. The on-lne adjustent of RBF neural network predctve odel Due to the error would be accuulated wth dynac characterstcs change, off lne RBF neural network can t eet the needs of the real-te predcton. In older to ake neural network predcton odel can real-te follow the change of the syste, t s necessary to buld on-lne update schee to adjust the paraeters of functon and the connecton weghts, to ensure the error between actual value and predcton trends to reduce. The goal of neural network adjustent on-lne s akng the predctve error as sall as possble. The object functon s followng
Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890 1 [ ( 1) ( 1)] J y k yp k (9) yk ( 1) Due to unable to get output of next step, t s reasonable to consder that the change range of actual syste output s slar to ts predcton, and obtan the followng forula The objectve functon can be splfed as : J( t) 1 [ y( k) y ( k)] y( k 1) y( k) y ( k 1) y ( k) (10) 5 p p p (11) The updatng of hdden layer nodes, the center locatons, the connecton weghts could be obtaned by gradent descent algorth. J c( k 1) c( k) c J ( k 1) ( k) (1) J ( k 1) ( k) Where paraeters s learnng rate of RBF center, s learnng rate of wdth, s learnng rate of weght. When output of neural network s,the predctve error s r( k ) xk ( ) x( k) c ( k) r( k) yk w ( k)exp( ) (13) ( k) The objectve functon of the neural network on-lne correcton could be adjusted as 1 1 x( k) c ( k) J y k w ( k)exp( ) 1 ( k) (14) To update the forula (1) (13) (14), the functon center, wdth, weght could be adjusted as forula (15) x( k) c ( k) x( k) c ( k) c( k 1) c( k) r( k) w ( k)exp( )( ) 1 ( k) ( k) x( k) c( k) x( k) c( k) ( k 1) ( k) r( k) w ( k)exp( )( ) 3 (15) 1 ( k) ( k) x( k) c ( k) ( k 1) ( k) r( k) exp( ) ( k) Updatng hdden layer nodes, the center locatons and the connecton weghts step by step, the neural network could follow the syste dynacally. 3. Sulaton analyss The cascade-rbf neural network can be sulated by MATLAB. Control odel can be descrbed as -180s a frst-order process wth te delay. The an loop s G 1 T () s 100s 1 e,the nor loop s G 0.5 -s P () s 3s 1 e,paraeters settng: node pont nuber of nput layer s 8, of the are output data
Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890 of syste,6of the are nput data of syste.; the nuber of plcaton layer s 10; the nuber of output layer s 1; predctve step s 10; learnng rate s 0.001; saplng perod s 10S; PID Paraeters turned n Table1, K p s proportonalty coeffcent, ntegral coeffcent, s dfferental coeffcent Table 1. PID Paraeters Paraeters K p Major Loop 0.4 0.0035 1 Mnor Loop 3 1 0 When the odel s atched, unt step response curve of cascade PID syste and cascade PID-RBF neural network syste were shown n Fg.5. as shown n Fg.5 and Table, copared wth cascade PID syste, the rse te, settng te, overshoot and steady state error of cascade-rbf neural syste are better than the forer. 0% and 10% step dsturbance were added respectvely to an loop and nor loop at the te of T=4000S and T=3000S.learnng fro Fg.5, step dsturbance alost have no effect on nor loop and settng te of cascade-rbf syste s shorter than ts cascade syste when an loop was effected by step dsturbance. K K K d K d Fg.5 Sulaton of the atched odel Table. The cascade control and cascade - RBF control perforance of the atched odel Control Schee Rse Te(s) Settng Te(s) Overshoot(%) Steady State Error Cascade 01 1330 16. 0 Cascade -RBF 110 540 5 0 When the odel of controlled plan has changed because of envronental change or the establshent of the forecast control odel s not accurate, the rato coeffcent, the te constant and lag te of syste all wll be changed. At that oent, the dynac perforances of systes were shown n Fg.6, Fg.7 and Fg.8. Fg.6 s cascade and cascade - RBF response curve of the rato coeffcent satch, the transfer -180s functon of controlled object n an loop s : G 1. T () s 100s 1 e ; Fg.7 s cascade and cascade - RBF response curve of the te constant satch, the transfer functon of controlled object n an -180s loop s : G 1 T () s 10s 1 e ; Fg.8 s cascade PID and cascade - RBF response curve of the lag -190s te satch, the transfer functon of controlled object n an loop s : G 1 T () s 100s 1 e. Learnng fro these fgures of odel satch, the overshoot of cascade-rbf neural network was less than cascade PID syste, the rse te and settng te of cascade-rbf neural network were uch better than cascade PID syste. 6
Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890 Fg.6 Cascade and cascade - RBF response curve of the rato coeffcent satch Fg.7 Cascade and cascade - RBF response curve of the te constant satch 4. Concluson Fg.8 Cascade PID and cascade - RBF response curve of the lag te satch 1. The predctve control schee that cobned RBF neural network and cascade PID could solve the bg lag of teperature control on S zorb heatng furnace, could control overshoot, the rse te and settng te effectvely. Its perforance of control s uch better than conventonal cascade PID control syste.. Copared wth cascade PID syste, the cascade-rbf neural network syste has better control character and stronger robustness to overcoe the dsturbance and odel satch on S zorb heatng furnace. References [1] Hou Xaong. S Zorb technology anual [M]. Bejng, Chna petrochecal press, 013, 5-9. [] Zhang JC, Lu Yunq, An Gaojun. Adsorpton desulfurzaton technology to produce clean ol products [J]. Journal of checal progress, 008,0(11):1834-1845. [3] Meng Xuan. catalytc gasolne adsorpton desulfurzaton technology research [D]. East Chna unversty of scence and technology, 01. [4] Lu Den. Petroleu checal ndustry autoatc control desgn anual [M]. Bejng, Checal ndustry press, 000:61-65 7
Internatonal Journal of Scence Vol.4 No. 017 ISSN: 1813-4890 [5] Wang Haoyu, Zhang Yunsheng, Zhang Guo. Iproved PID algorth of tubular furnace and ts applcaton n vrtual nstruent [J]. Proceedngs of the CSEE 009, 30(4): 51-54. [6] Du X, X Y, L S. Dstrbuted odel predctve control for large-scale systes[c] Aercan Control Conference, 001. [7] Wang Y, Boyd S. Fast odel predctve control usng onlne optzaton[j]. Control Systes Technology, IEEE Transactons on, 010, 18(): 67-78. [8] Zhan J, Ishda M. The ult-step predctve control of nonlnear SISO processes wth a neural odel predctve control (NMPC) ethod[j]. Coputers & Checal Engneerng, 1997, 1 ():01-10. [9] P. Aadaleesan, N. Mglan, R. Shara, Pr. SahaNonlnear syste dentfcaton usng Wener type Laguerre Wavelet network odel Checal Engneerng Scence, 63 (008), pp. 393 3941. [10] Zhang M, Wang X, Lu M. Adaptve PID control based on RBF neural network dentfcaton. In: Proceedngs of the 17th IEEE nternatonal conference on tools wth artfcal ntellgence. 005. [11] Zheng, E., Xong, J. Quad-rotor unanned helcopter control va novel robust ternal sldng ode controller and under-actuated syste sldng ode controller. 014 Optk 8