JP2002329187A - Method for estimating operating condition for petrochemical complex - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a petrochemical complex operating condition estimating method capable of quickly, stably and simply estimating the operating condition of a facility for giving the optimum value of an objective function in linear programming. SOLUTION: The estimating method is provided with a step S2 for acquiring learning data by calculating the yield of products corresponding to a different operating condition for the facility and generating a neutral network(NN) model, a step S7 for searching and acquiring an operating condition corresponding to a condition inputted from a linear programming model in which a prescribed objective function is set up by using the NN model, a step S8 for calculating the value of the objective function set up in the linear programming model on the basis of the obtained operating condition, a step S9 for judging whether the operating condition is an optimum operating condition or not from the obtained value of the objective function, and a step S11 for inputting a newly changed condition to the NN model when the operating condition is not the optimum operating condition.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used in a petroleum complex for producing petroleum products from a stock oil, and is obtained by a linear programming method based on operating conditions of a device constituting the petroleum complex obtained by simulation. The present invention also relates to a method for estimating operating conditions of a petroleum complex for estimating operating conditions of a device that gives an optimum material balance of petroleum products.

[0002]

2. Description of the Related Art In a process industry such as a petroleum complex that produces various petroleum products from a feedstock, a linear programming method (LP: Linear P
rogramming) is used. This linear programming is performed for the purpose of setting a constraint condition based on the operating conditions of the equipment constituting the raw material, the final product, and the petroleum complex, and finding the operating condition that gives the maximum value of the objective function.

However, the linear programming is performed on the assumption that both the objective function and the constraints are linear functions of variables, and the reaction temperature and the reaction temperature are obtained as in a reactor constituting a petroleum complex. When the product yield is in a nonlinear relationship, the linear programming cannot be used as it is. For this reason, conventionally, a nonlinear reaction simulator for obtaining these relationships in a reactor and a system for performing a linear programming (LP system) are linked to obtain an optimum material balance using a hill-climbing method. I was Specifically, the result of the initial solution obtained on the LP system is input to a nonlinear reaction simulator to obtain operating conditions, and LP operation is performed again based on the obtained operating conditions.
The system performs arithmetic processing to determine whether or not the obtained value of the objective function has converged.
The operating conditions changed on the P system are again input to the nonlinear reaction simulator and calculated, and the same procedure is repeated until the value of the objective function converges.

[0004]

However, such a conventional method has the following problems. That is, if the recursive calculation is performed by changing the operating conditions in the LP system, it is necessary to execute the calculation process by the nonlinear reaction simulator as many times as the number of the recursive calculations, resulting in a problem that the time required for the optimization calculation becomes enormous. Also, if the LP system is linked to multiple nonlinear reaction simulators,
The combination also increases the number of recursive calculations of the LP system, so that the time required for the optimization calculation is further increased.

Further, in order to perform convergence calculation for both the LP system and the nonlinear reaction simulator, a predetermined amount of allowable error is set as a criterion for determining whether or not convergence has occurred. For this reason, if the search width of the variable is reduced in order to search for a further peak near the top of the peak of the optimal value of the objective function, for example, for the value Y1 of the objective function in the operating condition X1, the operating condition X1 When there is a value Y2 of the objective function under the operating condition X2 where is slightly increased, Y2 may be smaller than the objective function value Y1. In this case, if the calculation by the non-linear reaction simulator is terminated within the allowable error, the recursive calculation of the LP system is repeated between the operating conditions X1 and X2. That is, since the objective function oscillates before and after a certain operating condition value, the optimization calculation does not converge, but is terminated in the middle of the calculation, and there is a problem that the study efficiency is deteriorated.

Further, in the conventional configuration in which the LP system and the nonlinear reaction simulator are directly linked, it is necessary to transmit an input necessary for the calculation of the nonlinear reaction simulator via the LP system. There is a problem in that data must be defined and stored, and complicated procedures such as addition by mixing oil types must be performed.

An object of the present invention is to provide a method for estimating an operating condition of a petroleum complex, which can quickly, stably, and easily estimate an operating condition of a device that gives an optimum value of an objective function in a linear programming method. To provide.

[0008]

According to the present invention, a procedure for searching for an optimum operating condition using a neural network model is interposed between a judgment of an optimum operating condition based on an objective function in an LP system and a calculation process in a reaction simulator. This aims to achieve the above object. Specifically, the method for estimating the operating conditions of a petroleum complex of the present invention is used in a petroleum complex for producing a petroleum product from a feedstock, and based on the operating conditions of a device constituting the petroleum complex obtained by simulation, A method for estimating an operating condition of a petroleum complex for estimating an operating condition of the apparatus which gives an optimum material balance of the feedstock and the petroleum product obtained by a linear programming method, wherein the product according to different operating conditions of the apparatus By calculating the yield of, learning data is obtained, based on a neural network model generation procedure for generating a neural network model, and a condition input from a linear programming model in which a predetermined objective function is set, Using a neural network, search for driving conditions according to the conditions And operating conditions acquisition procedure that Tokusuru,
Based on the operating conditions obtained in this operating condition acquisition procedure,
An objective function calculation procedure for calculating the value of the objective function set in the linear programming model, and an optimal operation for determining whether or not an optimal operating condition is obtained from the value of the objective function obtained in the objective function calculation procedure A condition determination step, and a change condition input step of changing the condition and inputting a new condition to the neural network model when it is determined that the condition is not the optimum operation condition in the optimum operation condition determination procedure. And

The present invention can be used on a system in which an LP system and a reaction simulator are linked,
The neural network model generation procedure, the operating condition acquisition procedure, and the change condition input procedure are preferably constructed in a reaction simulator. Further, in the neural network model, for example, assuming a multi-input / one-output neuron unit, operating conditions such as a reaction temperature and a pressure of the apparatus are set as inputs X1, X2, X3.
Is set, the connection weight at each input is W1,
When W2, W3,..., The relationship between the operating condition and the yield is given by a continuous function such as a sigmoid function by the following equation (1).

[0010]

(Equation 1)

Further, the objective function is given as a function of profit based on the material balance of the raw material and the petroleum product when the petroleum complex is operated, and can be obtained by the following equation (2), for example.

[0012]

(Equation 2)

In the formula (2), utilities represent electric power, steam, fuel, water, and the like, and auxiliary materials represent catalysts and chemicals used in the apparatus. Of these variables, the only independent variables are raw material usage and operating conditions, and prices are determined by market conditions. Product production, utility usage, auxiliary material usage, and the like are dependent variables that change depending on the selection of raw materials and the operating conditions of the apparatus.

According to the present invention, by providing a neural network model generation procedure and an operating condition acquisition procedure, a neural network model can be used based on conditions input from a linear programming model of an LP system. The operating condition according to the operating condition can be obtained in the operating condition obtaining procedure. Therefore, it is not necessary to obtain the operating condition of the device by simulation every time the recursive calculation is performed in accordance with the change of the condition of the LP system. And the speed of the optimization operation can be increased. Specifically, by performing the calculation using the LP system using such a neural network, the calculation time can be reduced by several thousandths compared to a conventional system in which the nonlinear reaction simulator and the LP system are directly linked. Can be shortened to

Further, by providing the neural network model as a sigmoid function of equation (1),
Since the relationship between the operating conditions and the yield can be unambiguously determined using a continuous function, the inversion of the calculated value of the objective function does not occur as in the case where the reaction simulator is repeatedly operated. A stable optimum value can be obtained by preventing the vibration phenomenon. Further, by providing the operating condition acquisition procedure, the relationship between the operating condition and the product yield can be established in the neural network model as compared with the case where the LP system and the reaction simulator are directly linked. There is no need to define and incorporate various data, and LP
The structure of the system is simplified.

In the above, the present invention is preferably employed when the relationship between the operating conditions of the apparatus and the product is given as a non-linear function. In the case of such a non-linear function, it is not possible to directly estimate the operating condition that gives the optimum material balance only by the linear programming, and the usefulness of the present invention is high.

Further, the above-mentioned apparatus is an ethylene production apparatus,
The present invention is preferably employed in any of a continuous catalyst regeneration type catalytic reformer, a vacuum gas oil desulfurizer, and a fluid catalytic cracker. Note that the method for estimating operating conditions of the present invention is not limited to being applied to only one of the devices constituting the petroleum complex, and may be applied to a plurality of devices in parallel. In such an apparatus, the relationship between the operating conditions and the yield is likely to be a non-linear function, and the linear programming method cannot be applied as it is. Therefore, the utility of adopting the present invention is high as described above.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows a petroleum complex according to an embodiment of the present invention, which is composed of crude oil, FRN (Full Range Naphtha), and LN.
(Light Naphtha) as a raw material oil, a system for producing petroleum products such as gasoline blends, kerosene blends, light oil blends, heavy A blends, and heavy B blends. , You can get these petroleum products.

The petroleum complex is composed of a plurality of fractionating devices, desulfurizing devices, manufacturing devices, etc. Specifically, the atmospheric distillation device 1, the vacuum distillation device 2, the vacuum gas oil desulfurization device 3, the fluid catalytic cracking device 4 , Methyl ethyl ketone production device 5, second hydrodesulfurization device 6, third hydrodesulfurization device 7, fourth hydrodesulfurization device 8, naphtha fractionation device 9, ethylene production device 1
0, naphtha hydrodesulfurization unit 11, continuous catalyst regeneration type catalytic reforming unit 12, pyrolysis gasoline hydrogenation unit 13, C5 hydrogenation unit 14, BTX (benzene / toluene / xylene) production unit 15, styrene monomer production unit 16, cyclohexane producing device 17, hydrogenated petroleum resin producing device 18, polybutene producing device 19, IP solvent producing device 20, para-xylene producing device 21, high deal device 22, reformate fractionating device 23, reformed gasoline distillation device 24, And a third aromatic extraction device 25.

The production of petroleum products in such a petroleum complex is carried out by appropriately blending the products produced by the above-described apparatuses 1 to 25. Taking the regular gasoline as a gasoline blend as an example, the regular gasoline is obtained by mixing the volatile oil fraction produced by the fluid catalytic cracking unit 4 with the continuous catalyst regeneration type catalytic reforming unit 12.
The main component is so-called reformate (reformed gasoline) produced by the above-mentioned method, and other base materials are blended in order to satisfy product specifications such as a distillation point and an octane number.

In such a petroleum complex, each device 1 that gives the most profitable and optimal material balance is provided.
An LP system is used to estimate ~ 25 operating conditions, and the operating conditions of each device are determined based on the optimal values obtained with this LP system. This LP system 3
When the operating conditions of the devices 1 to 25 fluctuate to 0,
The vacuum gas oil desulfurization unit 3, the fluid catalytic cracking unit 4, the ethylene production unit 10, the continuous catalyst regeneration type catalytic reforming unit 12, which greatly affects the product yield and in which the relationship between the operating conditions and the product yield is nonlinear. A non-linear reaction simulator 50 for simulating the relationship between the operating conditions and the product yield is connected. The LP system 30 and the non-linear reaction simulator 50 are stored and held as software in a computer installed in a petroleum complex, and when the operator starts the software, the LP system 30 and the nonlinear reaction simulator 50 are operated on an operation system for controlling the operation of the computer. Operate as a program to be expanded.

The LP system 30 includes an LP model setting section (LP model section) 31 for setting an LP model,
A Δ base modeling unit 32 for correcting a reference yield of a product obtained from the LP model set by the model setting unit 31. The LP system 30 is connected to the nonlinear reaction simulator 50 via the Δ base modeling unit 32. Have been. The LP model setting unit 31 includes crude oil, FRN, LN
This is a part for searching how much each petroleum product such as a gasoline blend or a light oil blend should be produced in order for the objective function to take the maximum value under the constraints set from the feedstock such as. Specifically, the LP model setting unit 31
Searches for the optimum material balance between the feedstock oil and the petroleum product by the continuous LP method to determine the yield of the product manufactured by each of the devices 1 to 25.

The Δ base modeling unit 32 is composed of crude oil, FR
This is a part for correcting the reference yield obtained by the LP model in consideration of the influence of the properties of the feedstock oil such as N and LN. For example, in the continuous catalyst regeneration type catalytic reformer 12, the following arithmetic processing is performed. Is performed. Now, in the continuous catalyst regeneration type catalytic reformer 12, the standard yields of gasoline (GAS), liquefied propane gas (LPG), and platform gasoline (PG) obtained as products are determined by the LP model setting unit 31 according to Table 1 below. And given the standard yield shown in This reference yield is treated as a fixed value in the LP system 30 without considering the change in the properties of the feedstock.

[0024]

[Table 1]

However, actually, each product GAS, LP
For G and PG, the actual yield varies depending on the properties of the feedstock oil. For example, naphthene content + 2 × aromatic content (N + 2A content) in the feedstock is selected as the feedstock property affecting the yield, It is assumed that the reference value is 51%, and that the N + 2A portion is 10% lower than the reference value due to the variation of the feedstock oil. At this time, in the Δ base modeling unit 32, a correction term for correcting the yield of each product when the N + 2A component is 10% lower is set to N + 2 in Table 1 above.
The correction yield is stored as the A correction, and the corrected yield is calculated by using the A correction so that the properties of the feedstock oil can be reflected on the product yield.

As shown in FIG. 2, the nonlinear reaction simulator 50 includes a reaction simulator body 51 and a neural network model setting section 52. The reaction simulator body 51 includes the target reaction device 3,
This is a part for simulating the relationship between the operating conditions of 4, 10, and 12, and the product yield when operated under the operating conditions. For example, assuming that the reaction in the reaction apparatus is governed by the Arrhenius type reaction rate equation, the reaction simulator body 51 incorporates the activation energy and frequency factor, which are the parameters of the reaction rate equation, and the amount of catalyst. (Diameter, height) etc. are also stored as data.

The main body 51 of the reaction simulator is as follows.
When reaction conditions such as a raw material supply amount, temperature and pressure, and a raw material composition are input as data, it is estimated what kind of reaction product is obtained by a catalyst under the reaction conditions.
That is, according to the reaction simulator body 51, for example,
How the reaction product changes when the reaction temperature is raised by 1 ° C,
It is possible to simulate what happens when the raw material composition changes.

Neural network model setting section 52
Is a part for learning a simulation result of the reaction simulator body 51 and setting a neural network model. As shown in FIG. 3, a neuron unit 521 set according to operable driving conditions is used as a unit. A neural network model is constructed by connecting a plurality of neuron units 521 according to the operating conditions of the reaction device.

The neuron unit 521 is a multi-input, one-output type unit.
When different reaction temperatures X1, X2, X3,... Are input and the connection weights W1, W2, W3,... Are input, a sigmoid function S representing the relationship between the reaction temperature X and the yield Y is output. . The neural network model set by such a neural network model setting unit 52 has a three-layer configuration including an input layer, a middle layer, and an output layer. The input layer is composed of a number of nodes according to operating conditions such as reaction temperature and feed amount, the intermediate layer is composed of 1 to 13, preferably 6 to 8 nodes, and the output layer is composed of an LP system. Is composed of a number of nodes corresponding to the restriction conditions to be processed. Input layer, middle layer,
And the nodes constituting the output layer are connected such that the output of each node constituting the input layer becomes the input of each node constituting the intermediate layer, and the output of each node constituting the intermediate layer becomes the input of the output layer. ing. Note that the number of nodes in the intermediate layer is appropriately determined according to the calculation accuracy. For such a neural network model setting unit 52, for example, a back propagation learning method is adopted, and learning is performed by changing a connection weight so that an error between an actual output and a desired output is minimized. Done.

Next, a method for estimating operating conditions of a petroleum complex by a system in which the above-described LP system 30 and the nonlinear reaction simulator 50 including the neural network model setting unit 52 are linked will be described with reference to the flowchart shown in FIG. Will be explained. (1) First, learning data is created using the reaction simulator body 51 (process S1). Specifically, data is created by changing input variables of certain operating conditions of the reaction simulator body 51. For example, if the reaction temperature is 490
The yield is calculated at each reaction temperature by changing the temperature from 1 to 510 ° C in steps of 1 ° C. The same applies to other operating conditions, and the reaction simulator 5
1 to generate learning data.

(2) The neural network model setting unit 52 creates a neural network model based on the created learning data (process S2: neural network model generation procedure). The neural network model is configured by combining a plurality of neuron units 521 set according to the number of driving conditions, and a known creation method can be used. (3) The accuracy of the created neural network model is evaluated (process S3). Specifically, the calculation is performed by comparing calculated values of the same input data. If the calculation accuracy is low, the input data is changed (process S4).
Improve accuracy.

(4) When the setting of the neural network model is completed, the LP model setting section 31 of the LP system 30 sets an LP model as an initial condition, and
The model is output to the neural network model setting unit 53 via the Δ base modeling unit 32 (processing S
5). (5) Based on the input LP model, the neural network model setting unit 52 searches for a combination of operating conditions according to the yield specified in the LP model (process S6), and performs an operation suitable for the LP model. The combination of conditions is output to the LP model setting unit 31 (process S7: operating condition acquisition procedure).

(6) Upon receiving the operating conditions, the LP model setting unit 31 calculates an objective function based on the operating conditions (process S8: objective function calculating procedure), and determines whether the value of the objective function has converged. That is, it is determined whether or not the operating conditions are optimal (process S9: optimal operating condition determining procedure). (7) If it is determined that the operating conditions are optimal, the actual operating conditions of the reactor are set based on the operating conditions, and the production is started (process S10). (8) On the other hand, if it is determined that the operating conditions are not optimal,
The condition of the LP model is changed and a new LP model is set (S11: change condition input procedure). The new LP model is processed again by the neural network model setting unit 53, and the same operation is repeated until optimal operating conditions are derived.

According to the present invention as described above, the following effects can be obtained. Neural network model generation procedure S
2. With the provision of the operating condition acquisition step S7, the operating condition according to the condition of the LP model can be obtained only by searching for the neural network model. Therefore, the simulation by the reaction simulator 51 is performed according to the input of the LP model. It is not necessary to perform the calculation, and the speed of the optimization calculation can be increased.

By the way, in the same personal computer, when the LP system 30 and the reaction simulator body 51 were directly connected to each other to calculate the product yield, it took 10.7 seconds to obtain the calculation result. When the operation processing of the reaction simulator body 51 is omitted using the neural network model, it takes only 8 milliseconds to obtain the operation result, and by adopting the present invention,
It can be seen that the calculation speed is dramatically improved.

Further, since the neural network model is given as a sigmoid function S which is a continuous function, the relationship between the operating conditions and the yield can be uniquely defined using the continuous function. The reverse of the calculated value of the objective function does not occur as in the case of the connection, the oscillation phenomenon is prevented, and a stable optimum value can be obtained. That is, as shown in FIG. 5, for example, in the relationship between the reaction temperature X and the yield Y, in the calculation of the optimum value by the conventional method in which the reaction simulator body 51 is directly connected, continuous calculation cannot be performed. The optimum value oscillates within a predetermined convergence width between the upper convergence value (□ in FIG. 5) and the lower convergence value (値 in FIG. 5). On the other hand, in the case of the sigmoid curve S, since Y for a predetermined reaction temperature X is uniquely determined,
Such a vibration phenomenon does not occur.

Further, by providing the operating condition acquisition procedure S7, it is possible to search for a combination of operating conditions using a neural network model based on the information on the minimum yield from the LP system 30. Therefore,
Compared with the conventional case of direct coordination, it is not necessary to define and incorporate various data in the modeling of the LP system, and the structure of the LP system 30 can be simplified.

The method for estimating the operating conditions according to the present embodiment is carried out by LP modeling of the vacuum gas oil desulfurization unit 3, the fluid catalytic cracking unit 4, the ethylene production unit 10, and the continuous catalyst regeneration type catalytic reforming unit 12. As a result, the nonlinear reaction simulation can be taken into the LP system 30 and the calculation can be performed at a high speed.

The present invention is not limited to the above embodiment, but includes the following modifications. In the above embodiment, the description of the neural network model has been made exclusively based on the relationship between the reaction temperature and the yield, but the present invention is not limited to this. That is, as the operating conditions of the present invention, the pressure of the reactor, the amount of catalyst, the catalyst life and the like may be adopted as parameters.
Any parameter that can be controlled and operated with respect to the yield of the device can be appropriately selected and incorporated into the neural network model.

In the above embodiment, the gasoline blend, the light kerosene blend, and the heavy oil blend are petroleum products, but the invention is not limited to this. That is, the present invention
It can be implemented for all petroleum products manufactured by the petroleum complex shown in FIG. 1, and is used, for example, when finding the optimal material balance of other petroleum products such as methyl ethyl ketone, ethylene, benzene, toluene, and xylene. You may.

Further, the petroleum complex in which the present invention can be implemented is not limited to the one shown in FIG. 1, and the present invention can be implemented in other types of petroleum complexes. In addition, specific structures, shapes, and the like when the present invention is implemented may be other structures and the like as long as the present invention can be implemented.

[0042]

According to the present invention as described above, the operating condition of the apparatus for giving the optimum value of the objective function in the linear programming is fast, stable, and stable even when a nonlinear reaction is included. It can be easily estimated.

[Brief description of the drawings]

FIG. 1 is a schematic view illustrating a structure of a petroleum complex according to an embodiment of the present invention.

FIG. 2 is a schematic diagram illustrating a structure of an LP system and a nonlinear reaction simulator in the embodiment.

FIG. 3 is a schematic diagram illustrating a configuration of a neuron unit configuring a neural network model in the embodiment.

FIG. 4 is a flowchart illustrating a procedure of a method for estimating an operating condition of a petroleum complex in the embodiment.

FIG. 5 is a graph comparing a sigmoid curve in the embodiment with a convergence value obtained by a method of directly connecting a linear programming system and a nonlinear reaction simulator.

[Explanation of symbols]

 S2 Neural network model generation procedure S7 Operating condition acquisition procedure S8 Objective function calculation procedure S9 Optimal operating condition determination procedure S11 Change condition input procedure

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) // C10G 9/00 C10G 9/00 11/18 11/18 35/24 35/24 45/72 45 / 72 (72) Inventor Kenzo Fujii 1-1, Shingu-cho, Tokuyama-shi, Yamaguchi F-term (reference) 4H029 AB03 AC14 BC06 BD08 5B056 BB92 HH00 5H004 GA30 GB02 KC13 KC22 KC27

Claims (4)

[Claims] 1. A feedstock and a petroleum product which are used in a petroleum complex for producing a petroleum product from a feedstock and which are obtained by a linear programming method based on operating conditions of a device constituting the petroleum complex obtained by simulation. A method for estimating operating conditions of a petroleum complex, which estimates operating conditions of the device that gives an optimal material balance, wherein learning data is obtained by calculating a product yield according to different operating conditions of the device. A neural network model generating procedure for generating a neural network model; and, based on conditions input from a linear programming model in which a predetermined objective function is set, operating conditions corresponding to the conditions using the neural network. Operating condition acquisition procedure for searching for and acquiring Based on the obtained operating conditions,
An objective function calculation procedure for calculating the value of the objective function set in the linear programming model; and optimal operation for determining whether or not the operating condition is optimal from the value of the objective function obtained in the objective function calculation procedure A condition determination step, and a change condition input step of changing the condition and inputting a new condition to the neural network model when it is determined that the condition is not the optimum operation condition in the optimum operation condition determination step. A method for estimating the operating conditions of a characteristic oil complex. 2. A method according to claim 1, wherein the relationship between the operating conditions of the apparatus and the yield of the product is given as a non-linear function. Condition estimation method. 3. The method for estimating operating conditions of a petroleum complex according to claim 1, wherein the neural network model generated in the neural network model generating procedure is represented by a continuous function such as a sigmoid function. A method for estimating the operating conditions of a petroleum complex. 4. The method for estimating the operating conditions of a petroleum complex according to claim 1, wherein the apparatus comprises: an ethylene production device, a continuous catalyst regeneration type catalytic reformer, which constitutes the petroleum complex; A method for estimating operating conditions of a petroleum complex, which is one of a vacuum gas oil desulfurization unit and a fluid catalytic cracking unit.