CN1226249C - Method for optimizing operation condition of xylene isomerization reactor - Google Patents

Abstract

The method for optimizing the operating conditions of the toluene isomerization reactor is to combine the radial basis function RBF network with multivariable interpolation and PLSR, that is, to apply linear regression and neural network methods in an integrated manner, adopt the structure of neural network RBFN, and use linear regression The method PLSR solution, the method of optimizing the operating conditions of the toluene isomerization reactor, the combination of the multivariable interpolation radial basis function RBF network and PLSR, and substituting the hidden layer output generated by each sample data into the above formula can form a formula with Similar to the multiple linear regression model, the PLSR method is used to solve this regression problem, and the ethylbenzene conversion rate, isomerization rate, C8 aromatics yield, and PX/∑X that reflect the reaction results of the reactor are selected as the model dependent variables, and the industrial reaction is used The actual production data of the reactor is used as a training sample to optimize the calculation of the reactor, find out the optimized operating parameters of the reactor, and make the dependent variable of the model optimal.

Description

Method for Optimizing Operating Conditions of Xylene Isomerization Reactor

1. Technical field

The invention relates to the optimization of process conditions for p-xylene isomerization reaction, in particular to a method for optimizing the operating conditions of a xylene isomerization reactor.

2. Background technology

The existing xylene isomerization reactor is an imported device and process. The device uses vacuum diesel oil, heavy coked diesel oil, light coked diesel oil, straight-run naphtha, and hydrogenated gasoline as raw materials to produce p-xylene, ortho xylene and benzene. The isomerization unit is based on the raw material of C8 aromatic hydrocarbon mixture containing poor o-xylene, m-xylene and ethylbenzene, which is converted into p-xylene concentration close to the equilibrium concentration under the action of hydrogen, catalyst and certain reaction temperature and reaction pressure. The mixture of C8 aromatic hydrocarbons can improve the yield of p-xylene. The existing isomerization unit adopts the ISOMAR process of U.S. Universal Oil Company (UOP). The design reaction pressure is 0.8-2.2Mpa, the temperature is 380-440℃, and the weight space velocity is 2.5-3.5hr ^-1 . Domestic SKI-400 catalyst. The isomerization product is a C ₈ A mixture with a concentration close to the equilibrium state. The light fraction generated by the reaction is removed through the deheptanizer, and then treated in the white earth tower to remove the unsaturated hydrocarbons generated by the reaction, and then sent to the xylene fractionation unit to remove C9 ⁺ A, then sent to the adsorption separation device.

The whole reaction process is a complex nonlinear process, in which the main reactions are:

(1) The isomerization of mixed xylenes is converted to different directions through the five-membered ring

(2) Isomerization of isobenzene

Side effects are:

(1) Disproportionation of xylene and ethylbenzene

(2) Hydrodealkylation of xylene and ethylbenzene

(3) Ring-opening cracking of ethylbenzene

(4) Paraffin hydrocracking

The main factors that affect the reaction are:

(1) Temperature Temperature is an important parameter of the reaction. The isomerization reaction is an exothermic reaction, but its thermal effect is very small, so the temperature has little overall effect on the chemical equilibrium. Since catalyst activity decreases as catalyst carbon deposits increase, the temperature must be increased to compensate. But the equilibrium concentration of C8 cycloalkane and p-xylene decreases with the increase of reaction temperature, so the pressure (hydrogen wind pressure) should be increased at the same time to achieve the desired effect.

(2) Pressure Pressure is an important parameter of the isomerization reaction. The conversion rate of ethylbenzene is achieved through the cycloalkane bridge, and the conversion of ethylbenzene is also the main feature of the isomerization reaction. The equilibrium concentration of naphthenes is related to hydrogen partial pressure and increases with increasing pressure. There are usually two ways to increase the partial pressure of hydrogen, increasing the total pressure or providing the concentration of hydrogen, so the pressure and hydrogen purity requirements of the initial and final stages of the reaction of this device are very different.

(3) Hydrogen-to-oil ratio Maintaining an appropriate hydrogen-to-oil ratio and hydrogen concentration has a certain impact on the conversion and yield of the isomerization reaction, and hydrogen can inhibit the carbon deposition of the catalyst.

(4) Space velocity (LHSV) Large space velocity is a feature of this device, up to 3.5~4.0h ^-1 , under certain other conditions, the conversion rate is low at high space velocity, and the conversion rate is improved at low space velocity. However, if the space velocity is too high, the side reactions will increase, the yield will be low and the catalyst will be easily deactivated.

The reaction system of the isomerization unit includes two series running at the same time: series 1 and series 2. The feed for isomerization is drawn from the side line of the raffinate tower of the adsorption separation unit, pretreated and mixed with the hydrogen from the circulating hydrogen compressor. The circulating hydrogen from 500# and 300# kettles are mixed together and sent into the heating furnace. After reaching the required temperature, they enter a series of reactors DC-701-1 and a second series of reactors DC-701-2. After the reaction, the two The series of products are brought together into the fractionation system.

The isomerization reactor is a fixed bed radial flow reactor, and the catalyst SKI-400 is filled in the reactor, and the loading capacity is 45,000 kg per series.

Xylene isomerization is the main way to produce p-xylene, and most of its production equipment has a relatively large scale. According to reports at home and abroad and our actual experience, large-scale modern petrochemical plants, due to a high degree of production standardization, generally have lower benefit potential than small-scale chemical plants. But even if there is a benefit of 1%, its absolute income increase is quite considerable, which has attracted the attention of many foreign experts, scholars and manufacturers, and there have been successful examples. However, almost all developers keep the key technologies strictly confidential, and their asking prices are often very expensive. Domestic work in this field is also underway, but relevant mature technical reports are still rare.

With the rapid development of computer technology and calculation, intelligence and control methods, the optimization operation of production equipment can also be carried out in many ways. It can be roughly divided into online and offline. The online method is currently based on advanced control theory, and the optimized control system designed can be directly debugged, controlled and optimized on the on-site production device. Correspondingly, hardware devices such as control systems must be equipped, and some need to establish corresponding models, so the cost is very high. The off-line method first needs to establish a model corresponding to the production device, and the modeling method can be divided into two categories. One is the kinetic model method. First, the corresponding mathematical model is established according to the chemical and physical mechanism of the production process, and then according to the The specific conditions of the device further determine the correction coefficient of the theoretical model (also known as the device factor), but the development process of this method is quite cumbersome and the cost is relatively high. In addition, the mechanism model is often appropriately simplified, and the theoretical model will inevitably deviate from the actual process, which will also affect the optimization effect.

Recently developed technologies such as neural networks, using these technologies and flexibly integrating them with statistical methods, can overcome the defects and deficiencies of purely applying statistical methods, and improve various performances of the built models, including fitting and Forecast ability, adaptability, robustness, etc. These modern stochastic methods will not fall into the local extremum region, especially suitable for network models, and can provide strong support for the integrated application of statistical and intelligent methods. The outstanding advantages of this type of optimization method are: according to the actual needs of the enterprise, advanced modern science and technology can be used to maximize the analysis and utilization of production operation data information, and the production potential of the device can be tapped without adding any equipment and raw materials. , improve the production efficiency of the device. In this sense, this is the cutting-edge technology and method most worthy of development and application.

From 1943 when psychologist W.S.McCulloch and mathematician W.Pitts studied and proposed the M-P neuron model, human research on neural networks has gone through more than half a century. Pioneering work by physicist Hopfield at Caltech. Later, the PDP (Parallel Distribution Processing) network thought of Rumelhart and McCelland and their research team played a role in fueling the arrival of a new upsurge in neural network research. In particular, the error back propagation (EBP) learning algorithm they proposed has become the most influential network learning method so far. The application of neural network has penetrated into various engineering fields. Because the training of neural network is a self-learning process, it is especially suitable for occasions where the rules and mechanism are not yet clear. In the field of chemical industry, neural network has been widely used in process modeling, pattern recognition, fault diagnosis and automatic control, etc., and has achieved gratifying results.

As we know, deriving various regression models based on statistical methods is another effective modeling tool. Compared with the modeling process of neural network (a gradual process that imitates the learning of human brain), the modeling based on statistical methods is a one-step process. If the network model is called "flexible model", then many regression models can be called "rigid model". Although the establishment of the regression model is fast and convenient, the effect of the model depends on the set model structure. Since there is no guarantee in advance what kind of model structure should be adopted, this brings great inconvenience to regression modeling. To this end, we propose methods that combine network and regression statistics so that the two complement each other.

Although there are some methods that can make the training process of BP network fall into a local minimum, it generally requires a huge amount of calculation, and the effect cannot be guaranteed. So we use the Radial Basis Function (RBF) network to fit the sample data. The RBF network not only has a good generalization ability, but also avoids cumbersome and lengthy calculations like the error backpropagation algorithm, so that the learning can be 10 ³ to 10 ⁴ times faster than the usual EBP algorithm.

Modeling methods have developed rapidly in recent years. From linear regression to artificial neural networks, and then to the combination of the two, scientists from all over the world are researching better and faster methods. Basically, linear regression is a basic method to establish an empirical model from actual sample data. The regression model is y=Xβ+s, and the least square method (LSR) is commonly used to estimate the value of parameter β. When there is a multicollinear relationship between the column vectors (ie, the respective variables) of the sample matrix X, the regression The values are very unstable and even have large deviations. Therefore, it is necessary to eliminate the multicollinear relationship between independent variables, and partial least squares regression (PLSR) is a better method to effectively eliminate the multicollinear relationship.

In complex nonlinear problems, the linear regression method is difficult to solve the problem perfectly. The artificial neural network provides a tool for nonlinear modeling with its powerful nonlinear expression ability. Radial Basis Function Network (RBFN) is a kind of artificial neural network with excellent performance. It not only has strong nonlinear data fitting ability, but also has good generalization ability for the constructed model.

Introduction to RBFN

The RBF network was proposed by Powell in 1985, and its essence is the radial basis function method of multivariate interpolation. The problem can be formulated as: given an n-dimensional point set { _xi } and the corresponding m-dimensional point set {y _i }, i=1, 2, ..., k, the interpolation problem is to ask for a function f(x) Make it meet the following interpolation conditions: f(x _i )=y _i

RBFN generally adopts a three-layer structure, including an input layer, a hidden layer, and an output layer, and the layers are fully connected, as shown in Figure 2.1. The input layer accepts input data and forwards it to each node in the hidden layer; the activation function of each node in the hidden layer is a radial basis function, and most of them use Gaussian functions, and their performance is related to the center parameter c and the width parameter σ. When the i-th input vector x _i is passed to the j-th hidden layer node, the processed output is:

${\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, c _{j and} σ _j are the center and width parameters of the node; in the output layer, only the output of the hidden layer is linearly weighted and used as the output of RBFN, and the rth output component is:

${\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; +</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; r\; j</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, w _rj is the connection weight, w _r0 is the bias item, and m is the number of nodes in the hidden layer.

Introduction to PLSR

Partial least squares regression is a new type of multivariate statistical data analysis method, which is called the second generation regression analysis method. It adopts information synthesis and screening technology in the regression modeling process, instead of directly considering the regression modeling of dependent variable set and independent variable set, but extracts several new comprehensive variables that have the best explanatory ability to the system from the variable system ( called PLS components, and are orthogonal to each other), and then use them for regression modeling. It differs from principal component regression in that principal component regression simply extracts comprehensive variables (called principal components) from the sample data X of the original variable, without considering the relationship with the dependent variable y. From the perspective of correlation, the extracted principal components do not necessarily contain enough information. However, PLSR not only extracts the PLS components from X, but also requires the covariance between the extracted PLS components and y to be the largest, and retains more correlation with the dependent variable, so that while eliminating the multicollinearity of the original variable, the established The regression model can still fully reflect the correlation between the independent variable and the dependent variable. The specific algorithm is attached as follows. For more information, please refer to the relevant literature.

PLSR algorithm:

PLS extracts mutually orthogonal components from the original data matrix, which not only retains more variance, but also retains more correlation with the dependent variable, so that while eliminating the multi-collinearity of the original variable, the established The regression model can fully reflect the correlation between independent variables and dependent variables.

In order to expand the scope of application, the multiple linear regression model takes its generalized form:

Y=XB+E

Among them, X is n×l (n>l) independent variable data matrix, Y is n×q (n>q) dependent variable data matrix, B is l×q parameter matrix, E is n×q residual difference matrix. Additionally, the columns of convention X and Y have been normalized.

PLSR usually uses the NIPALS algorithm, which decomposes the dependent variable data matrix Y while decomposing the independent variable data matrix, and tries to make the components extracted in X as close as possible to the components in Y (both are vectors in n-dimensional space ), that is, their correlation is as large as possible. The components extracted from X are often referred to as PLS components. Its algorithm steps are described as follows. Where s ₀ is an n-dimensional vector used to store intermediate data, h is an integer counter, δ ₁ and δ ₂ are two arbitrary small positive numbers given by the user to specify the precision.

(1) At the beginning of the program, the data of the independent variable and the dependent variable are sent to X and Y respectively, and the counter h is set to 1;

(2) A column vector y _i of optional matrix Y is sent into s ₀ as its initial value: y _i _s ₀ ;

The following steps are performed on the matrix X:

(3) Project the matrix X onto the column vector s ₀ : $x^T{\mathrm{the\; s}}_0/\left({\mathrm{the\; s}}_0^T{\mathrm{the\; s}}_0\right)\&DoubleRightArrow;u_h;$

(4) Normalize the vector u _h : u _h /‖u _h ‖_u _h ;

(5) Project the matrix X onto the row vector u _T ^h : $x{u}_h/u_h^T{u}_h\&DoubleRightArrow;t_h;$

The following steps are performed on the matrix Y:

(6) Project the matrix Y onto the column vector t _h : $Y^T{t}_h/t_h^T{t}_h\&DoubleRightArrow;v_h;$

(7) Normalize v _h : v _h /‖v _h ‖_v _h ;

(8) Project the matrix Y onto the row vector v _h : $Y{v}_h/v_h^T{v}_h\&DoubleRightArrow;{\mathrm{the\; s}}_h;$

(9) Check whether the column vector s ₀ is convergent: ‖ s ₀ -s _h ‖<δ ₁ ;

If not, then s _h _s ₀ , the program turns back to (3) to continue execution;

If so, the hth PLS component of the matrix X has been obtained so far, stored in t _h , and then the program proceeds to execute (10);

The following steps are to decompose the matrix X, Y:

(10) Calculate the load vector c _h of matrix X: $x^T{t}_h/t_h^T{t}_h\&DoubleRightArrow;c_h;$

(11) Regression of component s _h on t _h : $t_h^T{\mathrm{the\; s}}_h/t_h{t}_h\&DoubleRightArrow;b_h;$

(12) Remove the hth PLS component from the data matrix X: $x-t_h{c}_h^T\&DoubleRightArrow;x;$

(13) Remove the regression item from the data matrix Y: $Y-b_h{t}_h{v}_h^T\&DoubleRightArrow;Y;$

(14) Check whether all meaningful information in the data matrix X has been extracted: ‖X‖<δ ₂ ? ; If so, the program goes down (15);

If not, add 1 to the counter h, and verify that h<l;

If so, the program goes to (2) to continue execution; otherwise, the program proceeds to (15);

(15) The program ends.

3. Contents of the invention

The purpose of the invention is to provide a method of combining RBF (radial basis function) network and PLSR (partial least squares regression is a novel multivariate statistical data analysis method), that is, to apply linear regression and neural network method integratedly, The structure of the neural network RBFN is adopted, and the linear regression method PLSR is used to solve it, which avoids the difficulty of designing and training the RBFN, and is used to provide and control the optimization of the operating conditions of the xylene isomerization reactor.

Modeling of Xylene Isomerization Unit

Independent variable and dependent variable selection

The reaction system of the isomerization unit includes two series running at the same time. According to the influencing factors discussed above, the independent variable of the model is considered to be selected as:

(1) Catalyst age

(2) Two series of feed ethylbenzene content%

(3) Two series of feed m-xylene content%

(4) Two series of feed o-xylene content%

(5) Two series of reaction temperatures

(6) Two series of reaction pressures

(7) Two series of liquid hourly space velocity (LHSV)

(8) Hydrogen oil ratio of two series

Considering that many of the independent variables are calculated from more original data, these more original quantities can be used as independent variables in actual modeling to replace manual calculations, which facilitates calculations. These primitives are the independent variables mentioned above.

Select the model dependent variable as:

(1) Ethylbenzene conversion %

(2) Isomerization rate of p-xylene %

(3) C8A Recovery %

(4) PX/∑X in discharging

Modeling is the basis of the entire optimization system, and only when an accurate and reliable model is established can the next optimization work be meaningful. In the optimization system of xylene isomerization, we adopted the method of neural network modeling. This is an empirical modeling method. It does not need to consider the mechanism of the device when modeling. As long as the data obtained from the past operation of the device is used to train the network, the neuron network model of the device can be obtained. Both practice and theory have shown that neural network has high modeling accuracy and good forecasting ability. The RBF-PLS modeling method is adopted in the xylene isomerization optimization system, and the extracted component fraction in this method needs to be determined according to the actual object and data, so a cross-validation algorithm is also provided in the modeling module, Used to determine the most appropriate extraction fraction based on the sample data.

Optimization of operating conditions

The purpose of optimization is to determine the appropriate operating conditions to maximize the yield of certain products and at the same time minimize the cost of production, thereby improving the economic benefits of the enterprise. Optimization is to use stochastic/genetic algorithm under the model of neural network to search the independent variable space of operating conditions to find the optimal or better operating conditions.

The invention adopts the RBF-PLSR method to integrate the PLSR at the output end of the hidden layer of the RBFN. The number of hidden nodes of RBFN is taken as the number of training samples, that is, m=n, so that each hidden node corresponds to a training sample, and the central parameter c _i of the i-th node is taken as the i-th sample vector x _i . This design is equivalent to treating each sample point as a cluster center, and the width parameter σ _i can also be solved according to this idea. This solves the difficulties in RBFN structure design and main parameter selection.

The connection weight between the hidden layer and the output layer can still be determined by (2). Substituting the hidden layer output generated by each sample data into (2) formula can form a similar multiple linear regression model. Using the PLSR method to solve this regression problem can just solve the multicollinearity of independent variables that is easily caused by too small a sample size.

If the input and output data to be modeled are respectively {xi _} and { _yi } (ie independent variable and dependent variable), i=1, 2, . . . , k is the number of samples. Then the combination method of RBF and PLSR is as follows.

(1) Normalize x _i , i=1, 2, ..., k to [0, 1] (to eliminate the influence of dimension);

(2) Take all samples x _i as the radical base of RBF network, and then calculate the activation value of sample x _i relative to radical x _j according to the following formula, and form the input matrix A(k×k) of PLS;

(3) Establish the mapping relationship between matrices A and Y by using the above PLSR method.

The present invention integrates linear regression and artificial neural network methods to extract their strengths and supplement their respective deficiencies. RBF-PLSR adopts the structure of artificial neural network RBFN, and uses linear regression method PLSR to solve the problem, avoiding design and training. Due to the difficulty of RBFN, the built model has a concise analytical form, which is an excellent nonlinear modeling method. In this chapter, the RBF-PLSR method used to model the xylene isomerization unit is introduced.

4. Description of drawings

Figure 1 Overall structure of xylene isomerization optimization system

Figure 2 Workflow of xylene isomerization optimization system

Figure 3 is a schematic diagram of the RBFN network structure

5. Specific implementation

The xylene isomerization optimization system mainly establishes a neural network model based on the data obtained by the isomerization unit of the factory, and then, on the basis of this model, determines the optimal operating conditions for the production rate through optimization technology. The whole system can be divided into three parts: modeling, optimization and system maintenance, as shown in Figure 1:

The overall structural workflow of the xylene isomerization optimization system is shown in Figure 2: the main process parameters affecting the isomerization reaction

1 reaction temperature

The reaction temperature is an important parameter of the isomerization reaction. The isomerization reaction is an exothermic reaction, but its thermal effect is very small, so the overall effect of temperature on the chemical equilibrium is not great. Since catalyst activity decreases as catalyst carbon deposits increase, the temperature must be increased to compensate. But the equilibrium concentration of carbon 8 cycloalkane and p-xylene decreases with the increase of reaction temperature, so the pressure (hydrogen partial pressure) should be increased simultaneously to achieve the desired effect.

2 reaction pressure

The reaction pressure is also an important parameter of the isomerization reaction. The conversion rate of ethylbenzene is achieved through the cycloalkane bridge. The conversion of ethylbenzene is also the main feature of the isomerization reaction. The concentration of cycloalkane is related to the partial pressure of hydrogen and increases with the increase of pressure. There are usually two ways to increase the partial pressure of hydrogen, one is to increase the total pressure, and the other is to increase the concentration of hydrogen, so the pressure of the initial and final stages of the reaction of this device and the requirements for hydrogen purity are very different.

3 hydrogen oil ratio

Maintaining an appropriate hydrogen-to-oil ratio and hydrogen concentration has a certain impact on the conversion and yield of the isomerization reaction, and hydrogen can inhibit the carbon deposition of the catalyst.

4 airspeed

The liquid hourly space velocity (LHSV) is determined by the production unit and catalyst loading. Under certain other conditions, the conversion rate is low if the space velocity is large; while the conversion rate is high if the space velocity is small. However, if the space velocity is too high, the side reactions will increase, the yield will be low and the catalyst will be easily deactivated.

Optimization

According to the characteristics of the xylene isomerization reaction and the actual production situation of the industrial production plant, we select the feed composition of the reactor (that is, the content of ethylbenzene, o-xylene, m-xylene, p-xylene in the feed), Catalyst age, reactor capacity, reactor feed composition, reaction temperature, reaction pressure, hydrogen-to-oil ratio, and space velocity are the independent variables of the model, and the ethylbenzene conversion rate, isomerization rate, The yield of C8 aromatics and PX/∑X are model dependent variables. Using the actual production data of industrial reactors that meet certain conditions as training samples, the reactors are optimized and calculated to find the optimized operating parameters of the reactors, so that the model depends on The variable is optimal, thereby improving the production capacity of the reactor and the comprehensive economic benefits of the entire device.

Optimal Computing

According to the optimization method, we compiled an optimization software package for the xylene isomerization unit, and used the optimization calculation software to optimize the calculation of the operating conditions of the xylene isomerization reactor.

1 parameter setting

1.1 Setting Meter Parameters

The correction factors and ranges are shown in Table 1.

Table 1 Calibration 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 description is shown in Table 2

Table 2 Stream number description stream number illustrate stream number illustrate F7810 Raffinate F7912 recycle hydrogen F7811 Raffinate F7816 Hydrogen from 300# F7910 Raffinate F7817 Hydrogen from 500# F7911 Raffinate F7916 Hydrogen from 300# F6666 External supply of C8A feed F7917 Hydrogen from 500# F8001 DA702 bottom liquid F7032 DA702 top liquid F6039 DA603A raffinate oil F7815 FA701-1 discharge hydrogen F6218 DA603B raffinate oil F7915 FA701-2 discharge hydrogen F7812 recycle hydrogen F7031 FA702 gas to FG

1.2 Constraints

Calculate the data constraints Table 3.

Table 3 Calculation data constraints Balance rate range plus or minus 3% C8A Yield ≤99% Conversion rate of ethylbenzene >15%

1.3 Catalyst age

The starting time of the catalyst age is calculated according to the actual industrial data, that is, from January 4, 2000 to the end of April 2001.

2 Data generation

The data related to industrial reactors are processed and converted into data in a certain format for optimization calculations.

3 modeling

3.1 Cross Validation

Perform cross-validation on the data generated in 2, and select the number of extracted components with the lowest verification error of ethylbenzene conversion.

The verification results are shown in Table 4.

Table 4 Cross Validation Results number of training samples Validation sample number Minimal Extraction Fraction of Error 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

Modeling with cross-validated number of training samples and number of extracted samples.

4 Optimal calculation

4.1 Independent variable optimization parameter input

Input the compositions of feedstock F6039 and F6218, as shown in Table 5.

Table 5 feed composition F6039 F6218 EB% 13.80 13.80 MX% 57.24 56.35 0X% 18.75 17.66 PX% 1.39 1.25

4.2 Optimization parameter selection

Select independent variable optimization parameters: temperature, pressure, hydrogen-to-oil ratio, space velocity, and set the upper and lower limits for optimization. The setting results are shown in Table 6.

Table 6 The upper and lower limits of independent variable optimization parameters 1 series of upper and lower limits 2 series upper and lower limits temperature °C 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 airspeed 2.00 3.86 2.00 3.86

Select dependent variables to optimize parameters: conversion rate of ethylbenzene, isomerization rate, yield of C8 aromatics, PX/∑X.

4.3 Optimal calculation

4.3.1 Random search algorithm to optimize calculation results

The parameters of the random search algorithm are set as follows: the minimum search step size is 0.001; the number of randomly scattered points each time is 20. The optimization calculation results of different training sample numbers are shown in Table 7.

Table 7 Random search algorithm optimization calculation results number of training samples 60 70 74 80 90 100 extract fraction 7 7 6 8 7 11 1 series reaction temperature (℃) 400.6 401.0 400.5 401.0 400.0 401.6 2 series reaction temperature (℃) 399.1 399.5 399.0 399.9 398.7 400.7 1 series reaction pressure (Mpa) 1.27 1.29 1.28 1.30 1.27 1.32 2 series reaction pressure (Mpa) 1.27 1.29 1.27 1.30 1.27 1.31 1 series hydrogen oil ratio 8.60 8.24 8.20 7.68 8.03 7.47 2 series hydrogen oil ratio 7.11 6.79 6.81 6.32 6.68 6.17 1 series LHSV 2.54 2.66 2.65 2.80 2.69 2.85 2 series LHSV 2.52 2.64 2.62 2.76 2.65 2.79 Conversion rate 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 Calculation results optimized by differential evolution algorithm

The parameters of the differential evolution algorithm are set as follows: the number of groups is 100; CR is 0.8; F is 0.8; the maximum number of generations is 100; the algorithm strategy is 1. The calculation results of differential evolution algorithm optimization are shown in Table 8.

Table 8 Calculation results of differential evolution algorithm optimization number of training samples 60 70 74 80 90 100 extract fraction 7 7 6 8 7 11

1 series reaction temperature (℃) 400.6 401.0 400.5 401.0 400.2 401.6 2 series reaction temperature (℃) 399.1 399.5 399.0 400.0 398.7 400.7 1 series reaction pressure (Mpa) 1.27 1.29 1.28 1.30 1.27 1.32 2 series reaction pressure (Mpa) 1.27 1.29 1.27 1.30 1.27 1.31 1 series hydrogen oil ratio 8.60 8.24 8.20 7.68 8.03 7.47 2 series hydrogen oil ratio 7.11 6.79 6.81 6.32 6.68 6.17 1 series LHSV 2.54 2.66 2.65 2.80 2.69 2.85 2 series LHSV 2.52 2.64 2.62 2.76 2.65 2.79 Ethylbenzene conversion (%) 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 Enhanced differential evolution algorithm to optimize calculation results

The parameters of the enhanced differential evolution algorithm are set as follows: the number of groups is 100; CR is 0.8; F is 0.8; the maximum number of generations is 100; the algorithm strategy is 1. Table 9 shows the optimized calculation results of the enhanced differential evolution algorithm.

Table 9 Optimized calculation results of enhanced differential evolution algorithm number of training samples 60 70 74 80 90 100 extract fraction 7 7 6 8 7 11 1 series reaction temperature (℃) 400.6 401.0 400.5 401.0 400.2 401.6 2 series reaction temperature (℃) 399.1 399.5 399.0 400.0 398.7 400.7 1 series reaction pressure (Mpa) 1.27 1.29 1.28 1.30 1.27 1.32 2 series reaction pressure (Mpa) 1.27 1.29 1.27 1.30 1.26 1.31 1 series hydrogen oil ratio 8.60 8.23 8.20 7.68 8.04 7.48 2 series hydrogen oil ratio 7.11 6.79 6.81 6.32 6.68 6.18 1 series LHSV 2.54 2.67 2.65 2.80 2.69 2.85 2 series LHSV 2.52 2.64 2.62 2.76 2.65 2.80 Conversion rate 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 Optimization calculation analysis

From the optimization calculation process and the optimization calculation results in Table 6, Table 7, and Table 8, it can be seen that the optimization calculation results of the three optimization algorithms of random search, differential evolution, and enhanced differential evolution are not very different, indicating that the three optimization algorithms can be carried out Optimize calculations.

The actual optimization result has a great relationship with the number of training samples, that is, it has a great relationship with the actual industrial data involved in the calculation. The actual industrial production conditions are good, and the optimization calculation results are also good, otherwise the optimization calculation results are not ideal.

In actual optimization, different independent variable optimization parameters and dependent variable optimization parameters can be selected according to market conditions, product supply and demand conditions, and production device conditions, and the composition of feed F6039 and F6218 can also be optimized if necessary.

5 Conclusion

According to the optimization calculation, when the number of training samples is 74 and the number of extracted components is 6, the results of the independent variable and the dependent variable are optimal, that is, the reaction temperature of series 1 is 400.5°C, the reaction temperature of series 2 is 399.0°C, and the reaction pressure of series 1 is 1.28 Mpa, when the reaction pressure of series 2 is 1.27Mpa, the hydrogen-oil ratio of series 1 is 8.20, the ratio of hydrogen to oil of series 2 is 6.81, the LHSV of series 1 is 2.65, and the LHSV of series 2 is 2.62, the conversion rate of ethylbenzene is 27.84%, isomerization The yield of C8A is 16.93%, the yield of C8A is 95.48%, and the ratio of PX/∑X is 21.95%. The optimized ethylbenzene conversion rate is 2.7 percentage points higher than the industrial average ethylbenzene conversion rate of 25.14% in April.

(1) The specific formulas of hydrogen-to-oil ratio, liquid hourly space velocity, and equilibrium rate 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)

Among them, F7810, F7811, F7910, and F7911 are daily cumulative flow.

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

The formula for calculating catalyst age is as follows:

C _x A% _sxxy = (EB+MX+OX+PX)% _sxxy

v ₁ ＝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

in

Catalyst age at the end of L _i period (t raw material/kg catalyst)

A ₀ catalyst age at the beginning of the period (January 3, 2000) (t raw material/kg catalyst)

A Catalyst age increment (since the end of January 4, 2000) (t raw material/kg catalyst)

Cumulative value-added of raw materials treated by F _sum catalyst (since the end of January 4, 2000) (tons)

B Catalyst loading (45000kg/series) (kg)

F _day The amount of raw materials processed by the catalyst per day (tons)

F ₂ Cumulative flow meter value at 8:00 in recent days (tons)

F ₁ Accumulative flow meter value at 8:00 of the previous day (tons)

f flow correction coefficient (F7810, F7811, F7910, F7911)

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

Second series A ₀ ＝59.1544t raw material/kg catalyst

Claims (3)

1. The method for optimizing the operating conditions of the xylene isomerization reactor is characterized in that the radial basis function neural network of multivariable interpolation is combined with the linear regression method, that is, the structure of the radial basis function neural network is adopted, and the The linear regression method is solved, and the linear regression method is integrated into the output terminal of the hidden layer of the radial basis function neural network, and the number of hidden nodes of the radial basis function neural network is taken as the number of training samples, that is, m=n, so that each A hidden node corresponds to a training sample, and the central parameter c _i of the i-th node is taken as the i-th sample vector x _i ; all samples x _i are taken as the radial basis of the RBF network, and then obtained according to the following formula The activation value of the sample x _i relative to the radical x _j , and constitutes the input matrix A(k×k) of the linear regression; Substituting the hidden layer output generated by each sample data into the above formula can form a similar multiple linear regression model, use the linear regression method to solve this regression problem, and select the ethylbenzene conversion rate and isomerization rate that reflect the reaction results of the reactor , C8 aromatics yield, PX/∑X is the model dependent variable, using the actual production data of the industrial reactor as a training sample, optimize the calculation of the reactor, find out the optimized operating parameters of the reactor, and make the model dependent variable the most excellent, The arguments are: (1) Catalyst age (2) Two series of feed ethylbenzene content% (3) Two series of feed m-xylene content% (4) Two series of feed o-xylene content% (5) Two series of reaction temperatures (6) Two series of reaction pressures (7) Two series of liquid hourly space velocity (LHSV) (8) Two series of hydrogen-oil ratios. 2, by the method for the xylene isomerization reactor operation condition optimization described in claim 1, it is characterized in that radial basis function neural network RBFN adopts three-layer structure, comprises input layer, hidden layer and output layer, between layers In order to be fully connected, the input layer accepts input data and forwards it to each node in the hidden layer; the activation function of each node in the hidden layer is a radial basis function. 3. The method for optimizing the operating conditions of the xylene isomerization reactor according to claim 1 is characterized in that the optimization parameters are selected from the independent variable optimization parameters: temperature, pressure, hydrogen-oil ratio, space velocity, and the upper and lower limits of optimization are set : Upper and lower limits temperature °C 380 420 Pressure Mpa 1.2 1.4 Hydrogen oil ratio 4.0 9.0 airspeed 2.00 3.86

Images