CN1510018A - Method for optimizing operation condition of xylene isomerization reactor - Google Patents

Abstract

A method for optimizing the operating parameters of the xylene isomerizing reactor includes such steps as combining the radial basic function RBF network of multi-variable interpolation with PLSR, inserting the default output generated by various sample data in the equation to obtain a regression model, using PLSR method to find out the regression question, choosing the ethylbenzene transform rate, isomerizing rate, C8 arylhydrocarbon output rate, etc as the dependent variables of said model, using the actual data of industrial reactor as the training samples, and performing the calculation to find out the optimal parameters.

Description

The method that xylene isomerization reaction device operational condition is optimized

One, technical field

The present invention relates to the optimization of p-Xylol isomerization reaction processing condition, especially the method for xylene isomerization reaction device operational condition optimization.

Two, background technology

Existing xylene isomerization reaction device is device and the technology of introducing, and this device is a raw material with decompression diesel oil, heavy coker gas oil, light coker gas oil, virgin naphtha, hydrogenated gasoline, produces p-Xylol, o-Xylol and benzene.Isomerization unit is the carbon 8 aromatic hydrocarbons mixture raw materials to contain poor o-Xylol, m-xylene and ethylbenzene, facing under hydrogen, catalyst action and certain reaction temperature and the reaction pressure, change into carbon 8 aromatic hydrocarbons mixtures of p-Xylol concentration, improve the output of p-Xylol near equilibrium concentration.Existing isomerization unit is the ISOMAR technology that adopts UOP (UOP), and the design reaction pressure is 0.8～2.2Mpa, and temperature is 380～440 ℃, and weight space velocity is 2.5～3.5hr ^-1, transform the back and adopt homemade SKI-400 catalyzer.Isomerization product is the C near equilibrium state concentration ₈The A mixture removes the lighting end that dereaction generates through deheptanizer, handles through clay tower again, removes the unsaturated hydrocarbons that reaction generates, and is sent to the dimethylbenzene fractionation unit then, removes C9 ⁺A is sent to adsorption separation device again.

Entire reaction course is the non-linear process of a complexity, and wherein main reaction has:

(1) the xylol isomerization transforms to different directions through five-ring

(2) different benzene isomerization

Side reaction has:

(1) dimethylbenzene, ethylbenzene disproportionation

(2) dimethylbenzene, ethylbenzene hydro-dealkylation

(3) ethylbenzene open loop cracking

(4) paraffinic hydrocarbons hydrocracking

The factor of influence reaction mainly contains:

(1) the temperature temperature is the important parameter of reaction, and isomerization reaction is thermopositive reaction, but its heat effect is very little, so temperature is not very big to the total influence of chemical equilibrium.Because along with catalyst carbon deposition increases, catalyst activity reduces, and must improve temperature and compensate.But the equilibrium concentration of C8 cycloalkanes and p-Xylol reduces along with the raising of temperature of reaction, and pressurize (hydrogen blast) gets a desired effect so want simultaneously.

(2) pressure pressure is the important parameter of isomerization reaction, and the transformation efficiency of ethylbenzene reaches by the cycloalkanes bridge, and the conversion of ethylbenzene also is the principal feature of isomerization reaction.The naphthenic equilibrium concentration is relevant with the hydrogen dividing potential drop, and increases with the pressure raising.Improving the hydrogen dividing potential drop has two kinds of methods usually, improves stagnation pressure or the concentration of hydrogen is provided, so the requirement of the pressure in initial stage of this device reaction and latter stage and hydrogen purity is very different.

(3) hydrogen-oil ratio keeps appropriate hydrogen-oil ratio and hydrogen concentration that the transformation efficiency and the yield of isomerization reaction are had certain influence, and Hydrogen Energy suppresses the carbon distribution of catalyzer.

(4) air speed (LHSV) air speed is greatly characteristics of this device, reaches 3.5～4.0h ^-1, under the certain situation of other conditions, the big transformation efficiency of air speed is low, and the little transformation efficiency of air speed improves.But air speed is too big, and side reaction increases, the low and easy inactivation of catalyzer of yield.

The isomerization unit reactive system comprises two series of operation simultaneously: 1 series and 2 series, isomerized charging is to extract out from the raffinate tower side line of adsorption separation device, mix into process furnace through pre-treatment and with the hydrogen make-up of coming from circulating hydrogen compressor next recycle hydrogen, 500# and 300# still, enter series reaction device DC-701-1 and two serial reaction device DC-701-2 after reaching the requirement temperature, react latter two serial product and converge to and enter fractionating system together.

Isomerization reactor is a fixed bed radial flow reactor, loading catalyst SKI-400 in the reactor, and loadings is every serial 45000 kilograms.

Xylene isomerization is a main path of producing p-Xylol, and its production equipment has bigger scale mostly.According to report both domestic and external and our practical experience, large-scale modernization oil chemical plant installations, because production standardization degree is very high, the benefit potentiality are usually less than small-sized chemical plant installations.But even 1% benefit is arranged, it definitely increases income also is appreciable, thereby has caused the attention of external numerous expert, scholar and manufacturer, successful examples also occurred.But nearly all development person holds in close confidence key problem in technology, and its charge is often very expensive.Domestic work in this field is also launching, but relevant proven technique report is still rare.

Along with computer technology and calculating, intelligence and control method develop rapidly, the production equipment Optimizing operation also can be carried out in many ways.Can be divided into two kinds of online and off-lines haply.Online mode is at present based on advanced control theory, and designed optimal control system etc. can directly debug, control and optimize on the production equipment at the scene.Must be equipped with accordingly hardware devices such as Controlling System, what have also needs to set up corresponding model, so cost is very high.Offline mode at first needs to set up and the corresponding model of production equipment, the method of modeling can be divided into two classes again, one is the kinetic model method, at first set up corresponding mathematical model according to the chemistry and the Physical Mechanism of production process, and then further determine the correction coefficient (being referred to as the device factor again) etc. of this theoretical model according to the particular case of device, but the performance history of this method is quite loaded down with trivial details, and cost is also than higher.Mechanism model is often suitably simplified in addition, and theoretical model can depart from unavoidably to some extent to actual procedure, and these also all can influence optimization effect.

Technology such as the neural network of immediate development, adopt these technology, and with the integrated flexibly application of statistical method, can overcome the shortcomings and deficiencies of simple applied statistical method like this, and improved the various performances of institute's established model, comprise fitting to and prediction ability, adaptivity, robustness etc.These modern random devices can not be absorbed in the local extremum zone, especially can be applicable to network model, and the integrated application that can be statistics and intelligent method provides strong support.The outstanding advantage of this class optimization method is: can be according to the real needs of enterprise, the modern science and technology of uses advanced, the data message of analysis and utilization production operation to greatest extent, do not increasing under any equipment and the raw-material condition, the production potential of excavating gear, the productivity effect of raising device.On this meaning, this is the cutting edge technology and the method for the most worth exploitation utilization.

From nineteen forty-three psychologist W.S.McCulloch and mathematician W .Pitts research and propose the M-P neuron models and play today, human to the passed by course of over half a century of study of neural networks.The development work of the physicist Hopfield of California, USA Polytechnics.Thereafter, the PDP of Rumelhart and McCelland and research group thereof (Parallel Distribution Processing) network thought has then played the effect of adding fuel to the flames for the arrival of neural network research new upsurge.Especially their error anti-pass (EBP) learning algorithm of proposition becomes a kind of e-learning method that has the greatest impact so far.The application of neural network has been penetrated into each engineering field.Because the training of neural network is a self study process, so it is specially adapted to rule, the occasion that mechanism it be unclear that.In chemical field, neural network has been widely used in process model building, pattern recognition, trouble diagnosis and automatization control etc., and has obtained gratifying achievement.

FOR ALL WE KNOW, drawing various regression models based on statistical method is another kind of effectively modeling means.Compare with the modeling process of neural network (a kind of progressive process that imitates human brain study), then settle at one go based on the modeling of statistical method.If network model is called " flexible model ", so numerous regression models then can be described as " rigid model ".Though the foundation of regression model is quick and convenient, the effect of model depends on the model structure that sets.Owing to can't guarantee to adopt what kind of model structure in advance, this just brings very big inconvenience to regression modeling.For this reason, we propose network and return the method that statistics combines, so that both complementations.

Though certain methods is arranged, make the training process of BP net be absorbed in local minimum, but generally all need huge calculated amount, and effect can not guarantee.(Radial Basis Function, RBF) network comes the match sampled data so we adopt RBF.The RBF network not only has good popularization ability, and has avoided resembling loaded down with trivial details tediously long calculating the error back propagation algorithm, makes the study can be than common EBP algorithm fast 10 ³～10 ⁴Doubly.

The method development in recent years of modeling is very fast, from the linear regression to the artificial neural network, arrives both combinations again, and the various countries scientific worker is better in the research performance, faster method.Basic linear regression in fact is a basic skills of being set up empirical model by the actual sample data, and regression model is y=X β+ε, the value of least square method (LSR) estimated parameter β commonly used.When having again collinear relationship between the column vector (being each independent variable(s)) of sample matrix X, regressive

Be worth very unstablely, even produce very big deviation.Need to eliminate the multiple collinear relationship between independent variable(s), and partial least squares regression (PLSR) is the better method of the multiple collinear relationship of a kind of effective elimination at present for this reason.

In the nonlinear problem of complexity, linear regression method just is difficult to satisfactory having dealt with problems.And artificial neural network is with its strong non-linear ability to express, for Nonlinear Modeling provides an instrument.Footpath basis funciton net (RBFN) is a kind of artificial neural network of excellent property, and not only the nonlinear data capability of fitting is strong for it, and the model that makes up has good popularization ability.

The RBFN brief introduction

The RBF network proposed the RBF method of its essential multivariate interpolation by Powell in 1985.Problem can be illustrated as: given n dimension point set { x _iAnd corresponding with it m dimension point set { y _i, i=1,2 ..., k, interpolation problem require a function f (x) to make it to satisfy following interpolation condition exactly: f (x _i)=y _i

RBFN generally adopts three-decker, comprises input layer, hidden layer and output layer, and interlayer is for connecting fully, shown in Fig. 2 .1.Input layer is accepted the input data, and forward pass is given each node of hidden layer; The activation functions of each node of hidden layer is a RBF, and they adopt the Gaussian function mostly, and its performance is relevant with width parameter σ with Center Parameter c.As i input vector x _iReach j hidden layer node, be output as after the wherethrough reason:

$a_j=\exp (-||c_j-x_i{\left|\right|}^2/{\mathrm{\σ}}_l^2),---\left(1\right)$

C wherein _j, σ _lCenter and width parameter for this node; Output layer only the output of hidden layer is done linear weighted function and, and as the output of RBFN, r output component is:

$y_r=w_{r0}+\underset{j=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{rj}}{a}_j,---\left(2\right)$

W wherein _R1Be connection weight, w _R0Be bias term, m is the nodal point number of hidden layer.

The PLSR brief introduction

Partial least squares regression is a kind of novel multivariate statistics data analysing method, is called as s-generation regression analysis.It has adopted informix and triage techniques in the regression modeling process, it or not the regression modeling of considering that directly dependent variable rally and independent variable(s) are gathered, (be called the PLS composition but in variable system, extract some new generalized variables that system is had a best interpretations ability, and mutually orthogonal to each other), utilize their regression modelings then.What it was different with principal component regression is, principal component regression only merely extracts generalized variable (being called principal constituent) from the sampled data X of original variable, do not take into account the relation with dependent variable y, from the correlationship of dependent variable, the principal constituent of being extracted might not comprise abundant information.But PLSR not only extracts the PLS composition from X, also require the PLS composition extracted and the covariance of y to reach maximum, more and dependency dependent variable have been kept, thereby in the former variable multi-collinearity of cancellation, make the regression model of foundation still can reflect dependency between independent variable(s) and the dependent variable fully.It is as follows to attach the body algorithm, and more understandings can be with reference to relevant document.

The PLSR algorithm:

PLS extracts mutual orthogonal composition from original data matrix, they had both kept more variance, more and dependency dependent variable have also been kept, thereby in the former variable multi-collinearity of cancellation, make the regression model of foundation can reflect correlationship between independent variable(s) and the dependent variable fully.

In order to enlarge the scope of application, polynary linear regression model is got the form of its popularization:

Y＝XB+E

Wherein, X be n * l (the argument data matrix of n＞l), Y then be n * q (the dependent variable data matrix of n＞q), B is the parameter matrix of l * q, E is the residual matrix of n * q.In addition, each row of agreement X and Y are by stdn.

PLSR adopts the NIPALS algorithm usually, it is when decomposing the argument data matrix, also decomposing dependent variable data matrix Y, and managing to make the composition (both are the vector of n-dimensional space) among the as close as possible Y of the composition that extracts among the X, that is be that their dependency is big as far as possible.Often the composition that extracts is called the PLS composition from X.Its algorithm steps is described as follows.S wherein ₀For being used to deposit the n-dimensional vector of intermediate data, h is an integer counter, δ ₁, δ ₂Be two by the given arbitrarily small positive number of user, with specified accuracy.

(1) program begins, and independent variable(s) and dependent variable data are sent into X and Y respectively, and counter h is changed to 1;

(2) the column vector y of optional matrix Y _iSend into s ₀In, as its initial value: y _i s ₀

Following steps are that matrix X is handled:

(3) with matrix X projection in column vector s ₀On: $X^T{s}_0/\left(s_0^T{s}_0\right)\⇒u_h;$

(4) with vectorial u _hNormalization method: u _h/ ‖ u _h‖  u _h

(5) with matrix X projection in the row vectorial u _h ^TOn: ${\mathrm{Xu}}_h/u_h^T{u}_h\⇒t_h;$

Following steps are that matrix Y is handled:

(6) with matrix Y projection in column vector t _hOn: $Y^T{t}_h/t_h^T{t}_h\⇒v_h;$

(7) with v _hNormalization method: v _h/ ‖ v _h‖  v _h

(8) with matrix Y projection in the row vector v _hOn: ${\mathrm{Yv}}_h/v_h^T{v}_h\⇒s_h;$

(9) check column vector s ₀Whether restrain: ‖ s ₀-s _h‖＜δ ₁

If not, s then _h s ₀, program flow returns to (3) continue to carry out;

If, then so far obtained h the PLS composition of matrix X, be stored in t _hIn, program is down carried out (10) then;

Following steps are that Y decomposes to matrix X:

(10) load of compute matrix X vector c _h: $X^T{t}_h/t_h^T{t}_h\⇒c_h;$

(11) with ingredient s _hTo t _hReturn: $t_h^T{s}_h/t_h{t}_h\⇒b_h;$

(12) from data matrix X, remove h PLS composition: $X-t_h{c}_h^T\⇒X;$

(13) from data matrix Y, remove the recurrence item: $Y-b_h{t}_h{v}_h^T\⇒Y;$

(14) whether significant information all is extracted among the check data matrix X: ‖ X ‖＜δ ₂

If program is down carried out (15);

If not, then counter h adds 1, at check h＜l;

If program goes to (2) and continues to carry out; Otherwise program is down carried out (15); (15) EP (end of program).

Three, summary of the invention

The object of the invention provides the bonded method of a kind of RBF (RBF) net and PLSR (partial least squares regression is a kind of novel multivariate statistics data analysing method), promptly integrally use linear regression and neural net method, adopt the structure of neural network RBFN, find the solution with linear regression method PLSR again, avoided the difficulty of design and training RBFN, be used to provide and control xylene isomerization reaction device operational condition optimization.

The xylene isomerization unit modeling

Independent variable(s) and dependent variable are chosen

The isomerization unit reactive system comprises two series of operation simultaneously, according to the front influence factor is discussed, and considers that choosing the model independent variable(s) is:

(1) catalyzer agent age

The charging ethyl-benzene level % of (2) two series

The feed space xylene content % of (3) two series

The charging o-Xylol content % of (4) two series

The temperature of reaction of (5) two series

The reaction pressure of (6) two series

The liquid hourly space velocity (LHSV) of (7) two series

The hydrogen-oil ratio of (8) two series

Consider in the independent variable(s) that many amounts are by primary data computation more, in actual modeling can with these more the primary amount be used as independent variable(s) to substitute hand computation, made things convenient for calculating.These primary are exactly above-mentioned independent variable(s).

Choosing the model dependent variable is:

(1) conversion of ethylbenzene %

(2) p-Xylol isomerization rate %

(3) C8A rate of recovery %

(4) PX/ ∑ X in the discharging

Modeling is the basis of whole optimization system, has only and has set up model accurately and reliably, and next step optimization work is just meaningful.In the xylene isomerization optimization system, we have adopted the method for neural network modeling.This is a kind of experience modeling method, does not need to consider the mechanism of device when modeling, as long as the data that obtain when moving in the past with device are trained network, just can get the neural network model of auto levelizer.Practice and theory show that all neural network has higher modeling accuracy, and good prediction ability is arranged again simultaneously.In the employing of xylene isomerization optimization system is the RBF-PLS modeling method, being extracted into mark and need determining in this method according to the object and the data of reality, so a cross validation algorithm also is provided in MBM simultaneously, has been used for determining the optimum mark that is extracted into according to sampled data.

The optimization of operational condition

The purpose of optimizing is the suitable operational condition of decision, makes the productive rate of some product reach maximum, and it is minimum simultaneously production cost to be dropped to, thereby improves economic benefit of enterprises.Optimization is under the model of neural network, adopts at random/genetic algorithm, and the independent variable(s) space of operational condition is searched for, and seeks optimum or more excellent operational condition.

The present invention adopts the RBF-PLSR method PLSR to be integrated in the output terminal of RBFN hidden layer.The latent nodal point number of RBFN is taken as the number of learning sample, and promptly m=n makes each latent node corresponding with a learning sample, the Center Parameter c of i node _iJust be taken as i sample vector x _iDesign is equivalent to all regard each sample point as a cluster centre like this, also can find the solution width parameter σ by this thinking _lSolved the difficult point that RBFN structure design and significant parameter are chosen thus.

Connection weight between hidden layer and output layer still can be determined by (2) formula.Hidden layer output substitution (2) formula that each sampled data generated can be constituted one and similar multiple linear regression model.Use the PLSR method to find the solution this regression problem and just can solve the too little independent variable(s) multi-collinearity that easily causes of sample number.

If the inputoutput data of the modeling for the treatment of is respectively { x _iAnd { y _i(being independent variable(s) and dependent variable), and i=1,2 ..., k is a sample number.The combining method of RBF and PLSR is as follows so.

(1) with x _i, i=1,2 ..., k is normalized to [0,1] (to eliminate dimensional influence);

(2) with all sample x _iAs the footpath base of RBF net, then obtain sample x according to following formula _iWith respect to the basic x in footpath _jActivation number, and constitute the input matrix A (k * k) of PLS;

(3) the above-mentioned PLSR method of utilization is set up the mapping relations between matrix A and Y.

The present invention integrally uses linear regression and Artificial Neural Network, capture their strong point respectively, replenish deficiency separately, RBF-PLSR adopts the structure of artificial neural network RBFN, find the solution with linear regression method PLSR again, avoided the difficulty of design and training RBFN, the model of being built has simple and clear analytical form, a kind of good non-linear modeling method of can yet be regarded as.RBF-PLSR method that the modeling of brief introduction xylene isomerization unit is used in this chapter.

Four, description of drawings

The overall structure of Fig. 1 xylene isomerization optimization system

The workflow of Fig. 2 xylene isomerization optimization system

Fig. 3 is the RBFN schematic network structure

Five, embodiment

The xylene isomerization optimization system mainly is to set up neural network model according to the data that factory's isomerization unit has obtained, and on this model based, determines to make productive rate to reach optimum operational condition by optimisation technique then.Total system can be divided into modeling, optimizes and three parts of system maintenance, as shown in Figure 1:

The overall structure workflow of xylene isomerization optimization system, as shown in Figure 2:

Influence the main technologic parameters of isomerization reaction

1 temperature of reaction

Temperature of reaction is the important parameter of isomerization reaction.Isomerization reaction is thermopositive reaction, but its heat effect is very little, so temperature is not very big to the total influence of chemical equilibrium.Because along with catalyst carbon deposition increases, catalyst activity reduces, and must improve temperature and compensate.But the equilibrium concentration of carbon 8 cycloalkanes and p-Xylol reduces along with the raising of temperature of reaction, and pressurize (hydrogen dividing potential drop) gets a desired effect so want simultaneously.

2 reaction pressures

Reaction pressure also is the important parameter of isomerization reaction.The transformation efficiency of ethylbenzene is to reach by the cycloalkanes bridge, and the conversion of ethylbenzene also is the principal feature of isomerization reaction, and naphthenic concentration is relevant with the hydrogen dividing potential drop, and increases with the pressure raising.Improving the hydrogen dividing potential drop has two kinds of methods usually, and the one, improve stagnation pressure, the 2nd, the concentration of raising hydrogen is so the requirement of the pressure in initial stage of this device reaction and latter stage and hydrogen purity is very different.

3 hydrogen-oil ratios

Keep appropriate hydrogen-oil ratio and hydrogen concentration that the transformation efficiency and the yield of isomerization reaction are all had certain influence, and Hydrogen Energy suppress the carbon distribution of catalyzer.

4 air speeds

Liquid hourly space velocity (LHSV) is by production equipment and loaded catalyst decision.Under the certain situation of other condition, the big transformation efficiency of air speed is low; The little transformation efficiency height of air speed.But air speed is too big, and paying reaction increases the low and easy inactivation of catalyzer of yield.

Optimization method

According to the characteristics of xylene isomerization reaction and the practical condition of industrial installation, our feed composition of selected reactor (is an ethylbenzene in the charging, o-Xylol, m-xylene, the content of p-Xylol), catalyzer agent age, the reactor for treatment amount, reactor feed is formed, temperature of reaction, reaction pressure, hydrogen-oil ratio, air speed is the model independent variable(s), selected reflection reactor reaction result's conversion of ethylbenzene, isomerization rate, carbon 8 aromatics yields, PX/ ∑ X is the model dependent variable, utilization meets the industrial reactor actual production data of certain condition as learning sample, reactor is optimized calculating, find out the operating parameters of the reactor of optimization, make the model dependent variable for optimum, and then improve the throughput of reactor and the overall economic efficiency of whole device.

Computation optimization

We work out out xylene isomerization unit Optimization Software bag according to optimization method, utilize computation optimization software to carry out xylene isomerization reaction device operational condition computation optimization.

1 parameter setting

1.1 the strapping table parameter is set

Correction factor and range such as table 1.

Table 1 correction factor and range

Stream number	Correction factor	Range	Stream number	Correction factor	Range
						??F7810	??0.74116	??9999999	??F7912	?0.00046	??9999999
??F7811	??0.74116	??9999999	??F7816	?0.00021	??9999999
						??F7910	??0.74116	??9999999	??F7817	?0.00017	??9999999
??F7911	??0.74116	??9999999	??F7916	?0.00021	??9999999
						??F6666	??0.85501	??999999	??F7917	?0.00017	??9999999
??F8001	??0.71707	??9999999	??F7032	?0.64483	??999999
						??F6039	??0.7337	??9999999	??F7815	?0.00025	??9999999
??F6218	??0.73739	??9999999	??F7915	?0.00027	??9999999
						??F7812	??0.00044	??9999999	??F7031	?0.00138	??9999999

The stream number explanation is as table 2

The explanation of table 2 stream number

Stream number	Explanation	Stream number	Explanation
				F7810	Raffinate oil	F7912	Circulating hydrogen
F7811	Raffinate oil	F7816	The hydrogen that 300# comes
				F7910	Raffinate oil	F7817	The hydrogen that 500# comes
F7911	Raffinate oil	F7916	The hydrogen that 300# comes
				F6666	Outer for the C8A feed supplement	F7917	The hydrogen that 500# comes
F8001	Liquid at the bottom of the DA702 tower	F7032	The DA702 liquid of top of the tower
				F6039	DA603A raffinates oil	F7815	FA701-1 discharges hydrogen
F6218	DA603B raffinates oil	F7915	FA701-2 discharges hydrogen
				F7812	Circulating hydrogen	F7031	FA702 gas removes FG

1.2 constraint condition

Data calculated constraint condition table 3.

Table 3 data calculated constraint condition

The balance ratio scope	Positive and negative 3%
		The C8A yield	≤99％
Conversion of ethylbenzene	＞15％

1.3 catalyzer agent age

Catalyzer agent time of origin in age is by the Time Calculation of the actual industrial data of getting, promptly from January 4 to calendar year 2001 in 2000 by the end of April.

2 data generate

The industrial reactor related data is changed into after treatment the data that are used for computation optimization of certain format.

3 modelings

3.1 cross validation

2 data that generate are carried out cross validation, select conversion of ethylbenzene checking error minimum be extracted into umber.

Checking result such as table 4.

Table 4 cross validation result

Number of training	The checking sample number	The error minimum is extracted into mark
			????50	????1	????6
????60	????1	????7
			????70	????1	????7
????72	????1	????7
			????74	????1	????6
????76	????1	????8
			????80	????1	????8
????90	????1	????7
			????100	????1	????11
????150	????1	????19

3.2 modeling

Utilize the number of training and the extraction sample number of cross validation to carry out modeling.

4 computation optimization

4.1 independent variable(s) parameters optimization input

The composition of input charging F6039, F6218 is as table 5.

Table 5 feed composition

	????F6039	????F6218
			??EB％	????13.80	????13.80
??MX％	????57.24	????56.35
			??OX％	????18.75	????17.66
??PX％	????1.39	????1.25

4.2 parameters optimization is selected

Select the independent variable(s) parameters optimization: temperature, pressure, hydrogen-oil ratio, air speed, and set and optimize bound.Set result such as table 6.

Table 6 independent variable(s) parameters optimization bound

	1 serial bound		2 serial bounds
					Temperature ℃	????380	????420	????380	????420
Pressure Mpa	????1.2	????1.4	????1.2	????1.4
					Hydrogen-oil ratio	????4.0	????9.0	????4.0	????9.0
Air speed	????2.00	????3.86	????2.00	????3.86

Select the dependent variable parameters optimization: conversion of ethylbenzene, isomerization rate, carbon 8 aromatics yields, PX/ ∑ X.

4.3 computation optimization

4.3.1 random search algorithm optimization calculation result

The random search algorithm parameter is set at: the search minimum step is 0.001; Spreading at random counts at every turn is 20.The computation optimization result such as the table 7 of different number of training.

Table 7 random search algorithm optimization calculation result

Number of training	60	?70	?74	?80	?90	?100
							Be extracted into mark	7	?7	?6	?8	?7	?11
1 serial reaction temperature (℃)	400.6	?401.0	?400.5	?401.0	?400.0	?401.6
							2 serial reaction temperature (℃)	399.1	?399.5	?399.0	?399.9	?398.7	?400.7
1 serial reaction pressure (Mpa)	1.27	?1.29	?1.28	?1.30	?1.27	?1.32
							2 serial reaction pressure (Mpa)	1.27	?1.29	?1.27	?1.30	?1.27	?1.31
1 serial hydrogen-oil ratio	8.60	?8.24	?8.20	?7.68	?8.03	?7.47
							2 serial hydrogen-oil ratios	7.11	?6.79	?6.81	?6.32	?6.68	?6.17
1 serial LHSV	2.54	?2.66	?2.65	?2.80	?2.69	?2.85
							2 serial LHSV	2.52	?2.64	?2.62	?2.76	?2.65	?2.79
Conversion of ethylbenzene (%)	27.47	?27.06	?27.84	?27.68	?27.00	?26.36
							Isomerization rate (%)	16.68	?16.86	?16.93	?17.05	?16.89	?17.30
C8A yield (%)	94.78	?95.13	?95.48	?95.76	?95.21	?95.97
							?PX/∑X(％)	21.56	?21.69	?21.95	?21.89	?21.90	?21.75

4.3.2 differential evolution algorithm computation optimization result

The differential evolution algorithm parameter setting is: colony's number is 100; CR is 0.8; F is 0.8; Maximum algebraically is 100; Algorithm policy is 1.Differential evolution algorithm computation optimization result such as table 8.

Table 8 differential evolution algorithm computation optimization result

Number of training	60	?70	?74	?80	?90	?100
							Be extracted into mark	7	?7	?6	?8	?7	?11

1 serial reaction temperature (℃)	400.6	?401.0	?400.5	?401.0	?400.2	?401.6
							2 serial reaction temperature (℃)	399.1	?399.5	?399.0	?400.0	?398.7	?400.7
1 serial reaction pressure (Mpa)	1.27	?1.29	?1.28	?1.30	?1.27	?1.32
							2 serial reaction pressure (Mpa)	1.27	?1.29	?1.27	?1.30	?1.27	?1.31
1 serial hydrogen-oil ratio	8.60	?8.24	?8.20	?7.68	?8.03	?7.47
							2 serial hydrogen-oil ratios	7.11	?6.79	?6.81	?6.32	?6.68	?6.17
1 serial LHSV	2.54	?2.66	?2.65	?2.80	?2.69	?2.85
							2 serial LHSV	2.52	?2.64	?2.62	?2.76	?2.65	?2.79
Conversion of ethylbenzene (%)	27.47	?27.06	?27.83	?27.68	?27.00	?26.36
							Isomerization rate (%)	16.68	?16.86	?16.93	?17.05	?16.89	?17.30
C8A yield (%)	94.78	?95.13	?95.48	?95.76	?95.21	?95.97
							?PX/∑X(％)	21.56	?21.69	?21.95	?21.89	?21.90	?21.75

4.3.3 strengthen differential evolution algorithm computation optimization result

Strengthening the differential evolution algorithm parameter setting is: colony's number is 100; CR is 0.8; F is 0.8; Maximum algebraically is 100; Algorithm policy is 1.Strengthen differential evolution algorithm computation optimization result such as table 9.

Table 9 strengthens differential evolution algorithm computation optimization result

Number of training	60	?70	?74	?80	?90	?100
							Be extracted into mark	7	?7	?6	?8	?7	?11
1 serial reaction temperature (℃)	400.6	?401.0	?400.5	?401.0	?400.2	?401.6
							2 serial reaction temperature (℃)	399.1	?399.5	?399.0	?400.0	?398.7	?400.7
1 serial reaction pressure (Mpa)	1.27	?1.29	?1.28	?1.30	?1.27	?1.32
							2 serial reaction pressure (Mpa)	1.27	?1.29	?1.27	?1.30	?1.26	?1.31
1 serial hydrogen-oil ratio	8.60	?8.23	?8.20	?7.68	?8.04	?7.48
							2 serial hydrogen-oil ratios	7.11	?6.79	?6.81	?6.32	?6.68	?6.18
1 serial LHSV	2.54	?2.67	?2.65	?2.80	?2.69	?2.85
							2 serial LHSV	2.52	?2.64	?2.62	?2.76	?2.65	?2.80
Conversion of ethylbenzene (%)	27.47	?27.06	?27.83	?27.68	?27.01	?26.36
							Isomerization rate (%)	16.68	?16.86	?16.93	?17.05	?16.89	?17.30
C8A yield (%)	94.78	?95.13	?95.48	?95.76	?95.20	?95.97
							?PX/∑X(％)	21.56	?21.69	?21.95	?21.89	?21.90	?21.75

4.4 computation optimization analysis

By the computation optimization result of computation optimization process and table 6, table 7, table 8 as can be seen, it is little that three kinds of random searches, differential evolution, enhancing differential evolution are optimized optimization Algorithm calculation result difference, illustrate that three kinds of optimization algorithms all can be optimized calculating.

Actual optimum result and number of training have very big relation, promptly with the actual industrial data of participating in calculating very big relation are arranged, and industrial actual production situation is good, and its computation optimization result is also good, otherwise the computation optimization result is just not ideal.

Can select different independent variable(s) parameters optimization and dependent variable parameters optimization to be optimized according to the situation of market situation, product supply and demand situation, production equipment during actual optimization, can also the composition of charging F6039 and F6218 be optimized in case of necessity.

5 conclusions

According to computation optimization, in number of training is 74, being extracted into mark is 6 o'clock, independent variable(s) and dependent variable result optimal, promptly 1 serial reaction temperature is 400.5 ℃, 2 serial reaction temperature are 399.0 ℃, 1 serial reaction pressure is 1.28Mpa, and 2 serial reaction pressure are 1.27Mpa, and 1 serial hydrogen-oil ratio is 8.20,2 serial hydrogen-oil ratios are 6.81,1 serial LHSV is that 2.65,2 serial LHSV are 2.62 o'clock, and conversion of ethylbenzene is 27.84%, isomerization rate is 16.93%, the C8A yield is 95.48%, and PX/ ∑ X is 21.95%, and wherein the conversion of ethylbenzene of You Huaing is higher two 7 percentage points than industrial actual average conversion of ethylbenzene of April 25.14%.

(1) hydrogen-oil ratio, liquid hourly space velocity, the concrete formula of balance ratio are as follows:

I series H2/HC (mol/mol)=16.430139*F7812*H2% _S0705/ (F7810+F7811)

II series H2/HC (mol/mol)=13.922276*F7912*H2% _S0709/ (F7910+F7911)

I series LHSV=0.0007778166* (F7810+F7811)

II series LHSV=0.0007778166* (F7910+F7911)

Balance ratio=(F7810+F7811+F7910+F7911+F6666+F7816+F7817+F7916+F7917)/

(F8001+F7032+F7815+F7915+F7031)

Wherein F7810, F7811, F7910, F7911 are a day integrated flow.

(2) the dependent variable calculation formula is as follows:

Catalyzer agent calculation formula in age is as follows:

C ₈A％ _xxxx＝(EB+MX+OX+PX)％ _xxxx

v1＝F7810+F7910+F7811+F7911

L _i＝A ₀+A??????????????????(1)

A＝F _sum/B?????????????????(2)

F _sum＝∑F _day??????????????(3)

F _day＝(F ₂-F ₁)*f???????????(4)

L＝A ₀+∑((F ₂-F ₁)*f)/B

Wherein

L _iEnd of term catalyzer agent age (t raw material/kg catalyzer)

A ₀Initial agent age (on January 3rd, 2000) (t raw material/kg catalyzer) of catalyzer

Increment in A catalyzer agent age (from January 4th, 2000 to the end of term) (t raw material/kg catalyzer)

F _SumThe accumulative total increment of catalyst treatment material quantity (from January 4th, 2000 to the end of term) (ton)

B loaded catalyst (45000kg/ series) (kg)

F _DayCatalyzer is handled material quantity (ton) every day

F ₂Integrating flowmeter is 8:00 numerical value (ton) recently

F ₁Integrating flowmeter 8:00 day before yesterday numerical value (ton)

F flux modification coefficient (F7810, F7811, F7910, F7911)

A series of A ₀=59.1556 t raw material/kg catalyzer

Two series As ₀=59.1544 t raw material/kg catalyzer

Claims (3)

1, the method for toluene isomerization reactor operational condition optimization, the RBF RBF net that it is characterized in that multivariate interpolation combines with PLSR, promptly integrally use linear regression and neural net method, adopt the structure of neural network RBFN, find the solution with linear regression method PLSR again, PLSR is integrated in the output terminal of RBF net hidden layer, the latent nodal point number of RBFN is taken as the number of learning sample, be m=n, make each latent node corresponding, the Center Parameter c of i node with a learning sample _iJust be taken as i sample vector x _iWith all sample x _iAs the footpath base of RBF net, then obtain sample x according to following formula _iWith respect to the basic x in footpath _jActivation number, and constitute the input matrix A (k * k) of PLS;

The hidden layer output substitution following formula that each sampled data generated can be constituted one and similar multiple linear regression model, use the PLSR method to find the solution this regression problem, selected reflection reactor reaction result's conversion of ethylbenzene, isomerization rate, carbon 8 aromatics yields, PX/ ∑ X are the model dependent variable, utilize industrial reactor actual production data as learning sample, reactor is optimized calculating, find out the operating parameters of the reactor of optimization, make the model dependent variable for optimum

Independent variable(s) is:

(1) catalyzer agent age

The charging ethyl-benzene level % of (2) two series

The feed space xylene content % of (3) two series

The charging o-Xylol content % of (4) two series

The temperature of reaction of (5) two series

The reaction pressure of (6) two series

The liquid hourly space velocity (LHSV) of (7) two series

The hydrogen-oil ratio of (8) two series.

2, the method for optimizing by the described xylene isomerization reaction device of claim 1 operational condition, it is characterized in that RBFN adopts three-decker, comprise input layer, hidden layer and output layer, interlayer is for connecting fully, input layer is accepted the input data, and forward pass is given each node of hidden layer; The activation functions of each node of hidden layer is a RBF.

3, the method for being optimized by the described xylene isomerization reaction device of claim 1 operational condition is characterized in that parameters optimization selection independent variable(s) parameters optimization: temperature, pressure, hydrogen-oil ratio, air speed, and set and optimize bound: 1 serial bound 2 serial bounds Temperature ℃ ????380 ????420 ????380 ????420 Pressure Mpa ????1.2 ????1.4 ????1.2 ????1.4 Hydrogen-oil ratio ????4.0 ????9.0 ????4.0 ????9.0 Air speed ????2.00 ????3.86 ????2.00 ????3.86

Images