CN101620414A - Method for optimizing cracking depth of industrial ethane cracking furnace on line - Google Patents

Abstract

The invention relates to a method for optimizing the cracking depth of an industrial ethane cracking furnace on line. A cracking furnace yield neural network multi-model is established by utilizing a mechanism model and practical production data as training samples and combining a cracking raw material clustering result, model input variables comprise raw material density, sum of raw material linear chain alkane contents, ratio of raw material N-alkane content and isoparaffin content, feed flow, gas and hydrocarbon ratio, coil outlet temperature (COT) and coil outlet pressure, and model output variables comprise ethane yield and propylene yield (or triene total yield). A belonged submodel is judged according to a practical measuring value of an oil product, and the cracking depth output by the submodel is corrected on line by a cracked gas online analyzing instrument; an optimal cracking depth value and the gas and hydrocarbon ratio are confirmed by utilizing an SQP method and taking maximal economic benefits as a target through the constraints of the cracking depth and the gas and hydrocarbon ratio, and the cracking depth value is continuously optimized back and forth nearby the current cracking depth value of the cracking furnace. The method is reasonable, reliable, simple and easy, is easy to transplant and has wide adaptability.

Description

Method for optimizing cracking depth of industrial ethane cracking furnace on line

Technical field

The present invention relates to a kind of method for optimizing cracking depth of industrial ethane cracking furnace on line.

Background technology

Ethylene industry is the tap of petroleum chemical industry, and ethylene yield is a sign of weighing the petrochemical complex development level of a country.Pyrolysis furnace is in the leading position in the ethylene unit, the quality of its control is not only directly influenced the product quality and the output of whole ethylene producing device, but also influence the quiet run of downstream process units (as process units such as tygon, polypropylene, ethylene glycol).

Fig. 1 is typical pyrolysis furnace technological process and control chart, after hydrocarbon raw material F101 enters pyrolysis furnace BA101, earlier after the preliminary preheating of preheating section process, F102 mixes with dilution steam generation, further preheating and vaporization fully again, make its temperature be increased to the temperature that is lower than cracking reaction slightly, enter the reaction tube of pyrolysis furnace radiation section then, cracking reaction is mainly carried out in radiation section, it is thermonegative reaction, supplied with by bottom fuel gas F103 and sidewall fuel gas F104 burning in this part institute heat requirement, the hydrocarbon raw material in reaction tube heats up rapidly, cracking reaction takes place simultaneously generate ethene, propylene, butadiene, methane, products such as ethane.

Cracking severity is to weigh the important indicator that the interior cracking reaction of pyrolysis furnace is carried out degree, the principal element that influences cracking severity comprises COT (Coil Outlet Temperature, the pyrolysis furnace outlet temperature, hereinafter to be referred as COT), cracking stock composition, the residence time, vapour hydrocarbon ratio, wherein the COT temperature has the greatest impact to ethene, propene yield.Present domestic most of cracking of ethylene furnace apparatus all adopts COT to characterize cracking severity, and the COT setting value choose main with reference to pyrolysis furnace patent merchant provide at setting or the empirical value of design under the material condition, " cope with shifting events by sticking to a fundamental principle " often.Yet, it is complicated that domestic each ethylene production enterprise generally faces the cracking stock source, situations such as the oil property fluctuation is big, this makes the production run mode of fixation of C OT can not reflect in time that cracking severity and cracking product yield owing to cracking stock is formed and the variation of pyrolysis furnace operation conditions causes change; For this reason, East China University of Science takes the lead in developing and implements the cracking severity control system (patent No.: ZL200510025043.6), as shown in Figure 1, comprise the cascade control system that cracking severity control system and COT control system constitute, wherein the measured value of COT control system comes from the thermopair TI105 of cracking furnace tube outlet, its setting value is calculated and can be got by cracking severity controller AIC105 output, the cracking severity index is general adopt third second than (propylene and ethylene mass ratio) or first third than (methane propylene mass ratio), utilize on-the-spot pyrolysis gas to form in-line analyzer AI105 data as a reference, as the measured value of cracking severity controller AIC105.

The cracking severity control system changes the control mode that traditional industry field by using COT replaces cracking severity, but the setting value of this cracking severity control system is still determined by artificial at present, lack quantitative guidance, literature search finds still not have at present the industrial implementation precedent of cracking severity setting value on-line optimization, therefore, the raw material that the scene of making full use of possesses is formed information and pyrolysis gas in-line analyzer, binding data excavates and optimisation technique, and exploitation optimizing cracking depth of industrial ethane cracking furnace on line technology is significant.

Summary of the invention

The invention provides a kind of method for optimizing cracking depth of industrial ethane cracking furnace on line, cracking stock (is example with the naphtha) database is at first gathered and set up to the method, utilize FCM (Fuzzy C-Means, the fuzzy C-means clustering algorithm is hereinafter to be referred as FCM) obtain the oil product hierarchical cluster attribute center of N class oil product correspondence; Gather the actual industrial process data then the pyrolysis furnace mechanism model is carried out model tuning, utilize the mechanism model after proofreading and correct, and coverage is wider than the operating conditions data of on-the-spot operating mode commonly used, generate the data set of " raw material attribute-operating conditions-product yield " corresponding with N class oil product, and then the structure have the nonlinear Identification ability N neural network model, wherein, each sub neural network model input variable all is chosen for: material density, raw material straight-chain paraffin content sum, the ratio of raw material n-alkane isoparaffin content, feed rate, vapour hydrocarbon ratio, coil outlet temperature COT, boiler tube top hole pressure COP (Coil Outlet Pressure, the pyrolysis furnace top hole pressure, hereinafter to be referred as COP), the model output variable is chosen for: yield of ethene and propene yield; During field conduct, can measure property value according to new oil product, submodel under judging, and by this submodel calculating output pyrolysis product productive rate; Be predicated error and the uncertain disturbing factor that overcomes model, utilize the output of pyrolysis gas composition in-line analyzer that Neural Network model predictive is exported and constantly carry out on-line correction; Formulate the optimization aim function according to practical condition, market condition, use SQP (Sequential Quadratic Programming, Sequential Quadratic Programming method, hereinafter to be referred as: SQP) algorithm is optimized cracking depth index and vapour hydrocarbon ratio, make pyrolysis furnace guarantee to reach the maximum purpose of " diene " (being ethene, propylene mass yield sum) or " triolefin " (being ethene, propylene, butadiene mass yield sum) under the constant prerequisite of the cycle of operation.

Method for optimizing cracking depth of industrial ethane cracking furnace on line may further comprise the steps:

1. gather the historical floor data and the corresponding oil product attribute data of industrial ethylene pyrolysis furnace system;

2. choose the oil product attribute as the cluster attribute variable, carry out the oil product cluster: utilize the fuzzy C-means clustering algorithm that the oil product data are classified as the similar N class of crack characteristic;

3. according to the oil product cluster result of step in 2., utilize mechanism model to generate the data set of " raw material attribute-operating conditions-product yield " corresponding with N class oil product, N the neural network model that structure has the nonlinear Identification ability, wherein, mechanism model is the Simultaneous Equations that comprises cracking reaction kinetics equation, mass-conservation equation group, momentum conservation equation, energy conservation equation;

4. gather current working data and corresponding oil product attribute data,, calculate this oil product to the corresponding oil product distances of clustering centers of each submodel in conjunction with current oil product property value, the chosen distance minimum be current model: $d\left(i\right)=\sqrt{\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{|x_j-c_i^j|}^2},$ Selecting the minimum submodel of d (i) is current model, and wherein, the current measurement oil product of d (i) expression is to the corresponding oil product distances of clustering centers of each submodel, and the cluster centre of the corresponding oil product of each submodel is (c _i ¹, c _i ²... c _i ^j), current oil product property value is (x ₁, x ₂..., x _j), i=1 ... N, j=1 ... M, N are the oil product cluster classification number of step described in 2., and M is the cluster attribute variable number of step described in 2.;

5. for overcoming the uncertainty that model mismatch and external disturbance can bring to system usually, adopt keeping and on the forecast model basis of invariable error in future is made a prediction and carry out on-line correction: y _p(k+1)=y _m(k+1)+he (k), wherein, y _p(k+1) for next third second constantly of the k+1 after proofreading and correct than model predication value, y _m(k+1) be next third second constantly of k+1 than model prediction computation value, h is the error correction coefficient, e (k)=y (k)-y _mThird second of etching system is than actual value y (k) and model predication value y when (k) being k _m(k) error between;

6. according to the actual conditions of pyrolysis furnace, determine Optimizing operation variable and constraint condition, be target to the maximum with ethylene, propylene mass yield sum, predict that with the pyrolysis product yield behind the on-line correction neural network model is as the on-line optimization model, utilize the SQP method in conjunction with near the expert system rolling optimization cracking severity desired value of current C OT working point correspondence, obtain cracking severity index optimization net result.

The floor data of described step in 1. comprises the feed rate of pyrolysis furnace, steam flow, outlet temperature COT, top hole pressure COP, sidewall fuel gas flow, bottom fuel gas flow; The oil product attribute comprises the quality percentage composition of n-alkane, isoparaffin, alkene, naphthenic hydrocarbon, aromatic hydrocarbons, methane, ethene, propylene content number percent in the pyrolysis gas.

Described historical data acquisition range is historical floor data and the corresponding oil product attribute data in the period of 3 months to 2.

Described step 2. in, choose likening to of straight-chain paraffin total amount, n-alkane and isoparaffin and be the cluster attribute variable.

2. described step is by adopting the random initializtion cluster centre and repeatedly move the fuzzy C-means clustering algorithm, selecting the minimum corresponding result of cost function to realize the oil product data qualification.

The operating conditions that the data set of " the raw material attribute-operating conditions-product yield " of described step in 3. produces is 100%～120% of nominal situation scope basis.

The model of described step in 3. is the neural network regression model, wherein, each sub neural network model input variable all is chosen for: material density, raw material straight-chain paraffin content sum, the ratio of raw material n-alkane isoparaffin content, feed rate, vapour hydrocarbon ratio, coil outlet temperature COT, boiler tube top hole pressure COP, the model output variable is chosen for: yield of ethene and propene yield.

The error correction coefficient value scope of described step in 4. is [0.5,1.0].

The 6. middle cracking severity index optimization net result of described step has passed through amplitude limiting processing.

The rolling optimization of described step in 5., promptly, optimize performance index and only relate to from this following limited constantly time in each sampling instant, and to next sampling instant, this optimization period passes forward simultaneously, constantly the optimization performance index in this moment is optimized at each.

The invention provides a kind of method for optimizing cracking depth of industrial ethane cracking furnace on line, at first gather and construct cracking stock (is example with the naphtha) database, according to the oil product attribute, use the fuzzy C-means clustering algorithm that oil product is carried out cluster, obtain N class oil product and corresponding cluster centre; Secondly, utilize the data of ethane cracking furnace collection in worksite that the pyrolysis furnace mechanism model is proofreaied and correct, set up the multiple neural network model according to the oil product cluster result, each sub neural network model input variable all is chosen for: material density, raw material straight-chain paraffin content sum, the ratio of raw material n-alkane isoparaffin content, feed rate, vapour hydrocarbon ratio, coil outlet temperature COT, boiler tube top hole pressure, and the model output variable is chosen for: yield of ethene and propene yield; During field conduct, can measure property value according to new oil product, submodel under judging, and by this submodel calculating output pyrolysis product productive rate; Once more, utilize on-the-spot pyrolysis gas in-line analyzer that neural network model is carried out on-line correction, last, utilize the SQP method that the diene yield maximum of pyrolysis furnace is carried out tumbling-type optimization; The method makes full use of the data that pyrolysis gas in-line analyzer that present scene mostly possesses provides, utilize " forward prediction " of the regression model of mechanism model and historical data structure, be aided with " the reverse correction " of on-the-spot pyrolysis gas in-line analyzer, this method has at utmost reduced the dependence to the feed composition analyser, overcome the shortcoming that present scene does not mostly possess the feed composition analyser, in addition, the neural network multi-model process improves the precision of model, the on-line correction technology makes the model real-time update, the rolling optimization strategy improves the efficient and the on-line performance of system, native system changes the artificial blindness operator scheme of determining the cracking severity desired value of tradition, because method therefor is rationally reliable, simple, be easy to transplant, therefore have adaptability widely.

Description of drawings

Fig. 1 is the technological process control chart of pyrolysis furnace;

Fig. 2 is ethylene cracking severity on-line optimization overall plan figure;

The reference numeral key diagram:

Among Fig. 1, BA101 is a pyrolysis furnace, comprises the convection section of top and the radiation section of below; F101, F102, F103, F104 and F105 are respectively charging, steam, sidewall fuel gas, bottom fuel gas and pyrolysis gas; FIC101, FIC102, FIC103 and FIC104 are respectively feed rate meter, steam-flow meter, sidewall fuel gas flow meter, bottom fuel gas flow meter; AI101 is the material density meter, and TI105, AI105 are respectively pyrolysis furnace outlet pyrolysis gas temperature indicator and pyrolysis gas in-line analyzer, and TIC105 is the COT controller, and AIC105 is the cracking severity controller.

Embodiment

The present invention will be further described below in conjunction with drawings and Examples.

This specific embodiment is to being the optimization that the industrial ethylene pyrolysis furnace cracking of raw material is carried out with the naphtha.

The industrial crack furnace system of this instantiation as shown in Figure 1, BA101 is a pyrolysis furnace, comprise the top convection section and the radiation section of below; F101, F102, F103, F104 and F105 are respectively charging, steam, sidewall fuel gas, bottom fuel gas and pyrolysis gas; FIC101, FIC102, FIC103 and FIC104 are respectively feed rate meter, steam-flow meter, sidewall fuel gas flow meter, bottom fuel gas flow meter; AI101 is the material density meter, and TI105, AI105 are respectively pyrolysis furnace outlet pyrolysis gas temperature indicator and pyrolysis gas in-line analyzer, and TIC105 is the COT controller, and AIC105 is the cracking severity controller.Fig. 1 comprises the cascade control system that cracking severity control system and COT control system constitute, wherein the measured value of COT control system comes from the thermopair TI105 of cracking furnace tube outlet, its setting value is calculated and can be got by cracking severity controller AIC105 output, utilize on-the-spot pyrolysis gas to form in-line analyzer AI105 data as a reference, as the measured value of cracking severity controller AIC105.

This example is optimized cracking severity index (i.e. third second ratio) by the following method, the online setting value that provides of cracking severity controller AIC105 is provided, its solution as shown in Figure 2, overall controlling schemes comprises two levels, cracking severity Optimization Layer and cracking severity, vapour hydrocarbon compare key-course.

Cracking severity Optimization Layer operation mechanism is as follows: after the true-time operation data process data Coordination Treatment of pyrolysis furnace operation, input as depth optimization intelligent inference model, select affiliated submodel according to used oil product property measurement value, model is output as ethene, the propene yield of pyrolysis product, utilize on-the-spot on-line analysis instrument data to model output carrying out on-line correction again, adopt the SQP optimized Algorithm that the model after proofreading and correct is carried out rolling optimization, compare amplitude limit according to the third given second of scene again, carry out the amplitude limit constraint, export to the cracking severity key-course at last.

Cracking severity, vapour hydrocarbon adopt complicated PID control technology to carry out closed-loop control than the setting value of key-course from above-mentioned Optimization Layer acquisition controller.

The enforcement of cracking severity Optimization Layer may further comprise the steps:

1. gather and structure naphtha oil product database: gather the industrial ethylene pyrolysis furnace system service data in 3 months to 2 years in the past, comprise historical floor data and oil product attribute data.And current floor data and oil product attribute data gathered, as real time data.Wherein, floor data comprises the feed rate of pyrolysis furnace, steam flow, outlet temperature COT, top hole pressure COP, sidewall fuel gas flow, bottom fuel gas flow; The oil product attribute comprises the quality percentage composition of n-alkane, isoparaffin, alkene, naphthenic hydrocarbon, aromatic hydrocarbons, and methane, ethene, propylene etc. are formed percentage composition in the pyrolysis gas.

2. utilize the oil product data to carry out cluster, divide the oil product classification.The attribute that the naphtha database comprises has n-alkane, isoparaffin, alkene, naphthenic hydrocarbon and arene content, wherein n-alkane has the greatest impact to yield of ethene in the pyrolysis product, isoparaffin has the greatest impact to propene yield, n-alkane embodies the selectivity of yield of ethene with respect to propene yield with the ratio of isoparaffin, n-alkane and isoparaffin sum (being the straight-chain paraffin total amount) are to all gas-phase product yields and having the greatest impact, according to Analysis on Mechanism, choose the straight-chain paraffin total amount, n-alkane and isoparaffin liken attribute to into the naphtha oil product, utilize the fuzzy C Mean Method that naphtha oil product data are divided into N (N gets 6 in this method) class, think that the crack characteristic of oil product of inside of each class data is similar, can be similar to a neural network model and characterize, crack characteristic between the oil product of inhomogeneity data differs greatly, and is characterized by different neural network models.

FCM is described below to the fuzzy C-means clustering algorithm:

Sample data is made of the naphtha property value, and each group sample array is represented one group of naphtha attribute (straight-chain paraffin summation s ₁, n-alkane and isoparaffin ratio s ₂), to establish sample data and comprise n group naphtha attribute, each group naphtha attribute is x ₁={ s ₁, s ₂, at first set up n vectorial X={x ₁, x ₂..., x _n, represent n group naphtha attribute, establish this group sample data and have c cluster centre, then vectorial X={x ₁, x ₂..., x _nAnd c ambiguity group { X ₁, X ₂... X _cMembership U, by $X_i=\left\{(u_{x_i}\left(x_j\right),x_j)\right|x_j\∈X\}$ Embody, the essence of clustering algorithm is to make the cost function of non-similarity index reach minimum, and objective function is defined as:

$\left\{\begin{array}{c}J(U,C)=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{J}_i=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{\left|\right|x_j-c_i\left|\right|}^2\\ s.t{u}_{\mathrm{ij}}\&Element;\left[\mathrm{0,1}\right];\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}=1;0<\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}<n\end{array}\right.---\left(1\right)$

In the formula || x _j-c _i|| ²Be sample x _jTo ambiguity group X _i(the barycenter c of 1≤i≤c) _iDistance, can use d _IjRepresent.

And m ∈ in the formula [1, ∞) be a weighted index, m=2 in this algorithm.

Get d _IjBe Euclidean distance,, utilize Lagrangian multiplication to find the solution, constructed fuction J (1) formula:

$\stackrel{\&OverBar;}{J}(U,C,\mathrm{\&Lambda;})=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{d}_{\mathrm{ij}}^2+\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j(\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}-1)---\left(2\right)$

Here λ _j(1≤j≤n) is the Lagrange multiplier of n constraint formula of formula (1), the optimizing process of objective function J can be decomposed into 2 step-by-step optimizations and iterate and carry out.

1. U determines, optimizes C

Utilize formula (2) formula to c _iCarry out differentiate:

$\frac{\&PartialD;\stackrel{\&OverBar;}{J}}{\&PartialD;c_i}=\frac{\&PartialD;\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{d}_{\mathrm{ij}}^2}{\&PartialD;c_i}=\frac{\&PartialD;\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{(x_j-c_i)}^2}{\&PartialD;c_i}=\frac{\&PartialD;\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{(x_j-c_i)}^2}{\&PartialD;c_i}---\left(3\right)$

$=\frac{\&PartialD;\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m(x_j^2-2c_i{x}_j+c_i^2)}{\&PartialD;c_i}=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m(-2x_j+2c_i)$

Make formula (3) equal 0 and obtain a minimum necessary condition of reaching of formula (2):

$c_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{x}_j}{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m}---\left(4\right)$

2. C determines, optimizes U

Utilize formula (2) to λ _j, u _Ij(1≤i≤c) carry out differentiate gets:

$\frac{\&PartialD;\stackrel{\&OverBar;}{J}}{\&PartialD;{\mathrm{\&lambda;}}_j}=\frac{\&PartialD;\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j(\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}-1)}{\&PartialD;{\mathrm{\&lambda;}}_j}=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}-1---\left(5\right)$

$\frac{\&PartialD;\stackrel{\&OverBar;}{J}}{\&PartialD;u_{\mathrm{ij}}}=\frac{\&PartialD;[\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}^m{d}_{\mathrm{ij}}^2+{\mathrm{\&lambda;}}_j(\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}-1)]}{\&PartialD;u_{\mathrm{ij}}}=d_{\mathrm{ij}}^2\&times;{\mathrm{mu}}_{\mathrm{ij}}^{m-1}+{\mathrm{\&lambda;}}_j,(1\&le;i\&le;c)---\left(6\right)$

Make formula (5), (6) equal 0 respectively, obtain:

$\left\{\begin{array}{c}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{\mathrm{ij}}-1=0\\ d_{1j}^2\&times;{\mathrm{mu}}_{1j}^{m-1}+{\mathrm{\&lambda;}}_j=0\\ d_{2j}^2\&times;{\mathrm{mu}}_{2j}^{m-1}+{\mathrm{\&lambda;}}_j=0\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ d_{\mathrm{cj}}^2\&times;{\mathrm{mu}}_{\mathrm{cj}}^{m-1}+{\mathrm{\&lambda;}}_j=0\end{array}\right.---\left(7\right)$

Separate formula (7) and obtain a minimum necessary condition of reaching of formula (2):

$u_{\mathrm{ij}}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\&Sigma;}}}{\left(\frac{d_{\mathrm{ij}}}{d_{\mathrm{kj}}}\right)}^{2/(m-1)}}---\left(8\right)$

Above-mentioned two necessary conditions have constituted a simple iterative process of FCM algorithm, and algorithm steps is as follows:

(1) with value 0,1 random number initialization degree of membership matrix U, make it satisfy constraint condition in the formula (1).

(2) calculate c weighting cluster centre c with formula (4) _i(1＜i＜c).

(3) according to formula (1) if calculating target function. its relative last time target function value the change amount less than certain threshold value, the threshold value value of present embodiment is 0.05%, then algorithm stops.

(4) calculate new matrix U with formula (8), return step 2.

Above-mentioned algorithm also can first initialization cluster centre, and then execution iterative process, owing to can not guarantee that FCM converges on an optimum solution, the performance of algorithm depends on initial cluster center, therefore, can adopt the random initializtion cluster centre and repeatedly move FCM, select the minimum corresponding result of cost function.

Through above-mentioned cluster calculation, obtain optimum clusters number c=N=6, and the cluster centre corresponding with it, i.e. barycenter c _i, i=1 ... 6.

3. utilize the historical data of collection in worksite to input to mechanism model, according to the deviation of model output with the real process data, constantly the adjustment model parameter is proofreaied and correct the pyrolysis furnace mechanism model, and is consistent with real process output up to model output; Use the mechanism model after proofreading and correct to produce the wider set of data samples of covering operating mode scope, the structure training data is used to set up the neural network model of pyrolysis product yield prediction:

Utilize the existing raw material attribute data collection of gathering, to each class property of raw material, according to on-the-spot actual condition, in 120% scope to the nominal situation scope basis of material flow, vapour hydrocarbon ratio, outlet temperature, four variablees of top hole pressure, utilize mechanism model to produce the sample data that is used to set up neural network model.Herein, mechanism model is the Simultaneous Equations that comprises cracking reaction kinetics equation, mass-conservation equation group, momentum conservation equation, energy conservation equation.

To each class oil product attribute corresponding sample data set, the picked at random sample data 80% as training sample data collection, choose residue 20% as the forecast sample data set, in order to eliminate the influence of dimension separately, carry out normalized; The variable of process normalized enters neural network model and carries out computing, wherein the definite of weights coefficient obtained sample learning by network in each node, result after the network operations calculates the precision of prediction that mean square deviation is used to weigh neural network through after the anti-normalization with actual output data contrast.

At each class oil product attribute corresponding sample data set is set up sub neural network, each sub neural network model structure is the feedforward neural network of seven inputs, two outputs, the suitable method of gathering of definite employing of the hidden layer node number of sub neural network determines that the input variable of each sub neural network is:

(1) material density

(2) raw material straight-chain paraffin content sum

(3) ratio of raw material n-alkane isoparaffin content

(4) feed rate

(5) vapour hydrocarbon ratio

(6) coil outlet temperature

(7) boiler tube top hole pressure

The output variable of each sub neural network is:

(1) ethene mass yield

(2) propylene mass yield

The ethene of actual pyrolysis furnace, productivity of propylene and outlet temperature, oil product is formed, feed rate, vapour hydrocarbon ratio, cross-over temperature, the residence time, along the temperature of boiler tube pipe range, pressure distribution etc. direct relation is arranged all, but because the boiler tube tube wall heat flux distribution of actual pyrolysis furnace is unknown and constantly variation in the pyrolysis furnace operational process, any model all can only be similar to this nonlinear relationship that characterizes ethene, propylene and cracking operating mode and raw material, suppose that the intelligent yield model of setting up can be similar to this relation of sign, then its corresponding mathematical model is:

$\hat{E}(k+1)=f_e\{\mathrm{density}\left(k\right),\mathrm{sp}\left(k\right),\mathrm{nir}\left(k\right),\mathrm{ff}\left(k\right),\mathrm{sdr}\left(k\right),\cot \left(k\right),\mathrm{cop}\left(k\right)\}$ (9)

$\hat{P}(k+1)=f_p\{\mathrm{density}\left(k\right),\mathrm{sp}\left(k\right),\mathrm{nir}\left(k\right),\mathrm{ff}\left(k\right),\mathrm{sdr}\left(k\right),\cot \left(k\right),\mathrm{cop}\left(k\right)\}$

Wherein k represents current time, and k+1 represents next constantly, and the time interval is an optimization cycle,

Be respectively ethene, propene yield, density characterizes material density, and sp characterizes raw material straight-chain paraffin total amount, nir represents the ratio of raw material n-alkane isoparaffin content, and ff is a material flow, and sdr is a vapour hydrocarbon ratio, cot is a coil outlet temperature, and cop is the boiler tube top hole pressure.

At the actual industrial scene, raw material component characteristic, feed rate, top hole pressure belong to can not manipulated variable, is that cracking severity index propylene and ethylene mass ratio per and vapour hydrocarbon compare sdr but the model manipulated variable is the model independent variable of this method.

4. gather current operation floor data and naphtha and analyze data, at first utilize naphtha to analyze data chooser model, utilize step 3 to set up good submodel prediction ethene, propene yield again.

At first, the distances of clustering centers according to the current oil product property measurement value ratio of isoparaffin (the straight-chain paraffin summation, n-alkane with) and each sub neural network calculates and the immediate submodel of current oil product:

If the cluster centre of the corresponding oil product of each submodel is (c _i ^Sp, c _i ^Nir), i=1 ... N, current oil product property value is (x ₁, x ₂), then this oil product is as follows to the corresponding oil product distances of clustering centers of each submodel computing formula:

$d\left(i\right)=\sqrt{{|x_1-c_i^{\mathrm{sp}}|}^2+{|x_2-c_i^{\mathrm{nir}}|}^2}---\left(10\right)$

Then, utilize this submodel to calculate yield of ethene and propene yield, and then obtain cracking severity index third second ratio of Model Calculation.

5. for overcoming the uncertainty that model mismatch and external disturbance can bring to system usually, adopt following model on-line correction method:

y _p(k+1)＝y _m(k+1)+h·e(k) (11)

Y wherein _p(k+1) for next third second constantly after proofreading and correct than model predication value, y _m(k+1) be next third second constantly than model prediction computation value, h is the error correction coefficient, adjusts according to the effect of practical application, the present embodiment span is [0.5,1.0]; E (k)=y (k)-y _m(k) be that third second of current time system is than actual value y (k) and model predication value y _m(k) error between.

6. determine optimization aim, Optimizing operation variable and constraint condition; Utilize SQP to adjust optimum cracking severity index third second ratio, and adjust the vapour hydrocarbon ratio of pyrolysis furnace in conjunction with expert judgments, up to reaching optimization aim:

Suppose that the ratio that market ethene price, propylene price account for ethene, propylene price sum respectively is α and β, then the task of depth optimization promptly be seek the optimum cracking severity index third second ratio (with the vapour hydrocarbon than), make in ethylene, propylene weighted sum maximum:

Objective function: $J(K+1)=\max \{\mathrm{\&alpha;}\&CenterDot;\hat{E}(K+1)+\mathrm{\&beta;}\&CenterDot;\hat{P}(K+1)\}---\left(12\right)$

Constraint condition: $\left\{\begin{array}{c}\hat{E}(k+1)=f_e\{\mathrm{density}\left(k\right),\mathrm{sp}\left(k\right),\mathrm{nir}\left(k\right),\mathrm{ff}\left(k\right),\mathrm{sdr}\left(k\right),\cot \left(k\right),\mathrm{cop}\left(k\right)\}\\ \hat{P}(k+1)=f_p(\mathrm{density}\left(k\right),\mathrm{sp}\left(k\right),\mathrm{nir}\left(k\right),\mathrm{ff}\left(k\right),\mathrm{sdr}\left(k\right),\cot \left(k\right),\mathrm{cop}\left(k\right))\\ \mathrm{per}(k+1)=\hat{P}(k+1)/\hat{E}(k+1)\\ \mathrm{perlolm}<\mathrm{per}(k+1)<\mathrm{perhilm}\end{array}\right.---\left(13\right)$

Preceding two equatioies of this constraint condition have embodied the mutual relationship of pyrolysis furnace ethylene yield and productivity of propylene and raw material composition and process conditions, the 3rd is cracking severity index third second ratio, the 4th is inequality, embodied the variation range of optimization variable cracking severity index-third second ratio, this variation range is determined according to the adjustable extent of COT, and adjustable definite need of COT are determined according to scene pipe surface temperature temperature measurement data every day, perhaps require to determine according to the technologist, this constraint is guaranteed the security of cracking severity optimal control system on the one hand, and the cycle of operation of pyrolysis furnace is guaranteed.

Above-mentioned optimization problem is carried out simple global optimization can't adapt to the influence that uncertain factors such as oil property variation are brought.Therefore, we have introduced the thought of rolling optimization, promptly in the certain hour window, maximize the economic performance index, simultaneously in time adjust optimisation strategy according to the variation of actual production situation, by in the time window that progressively moves, constantly revising the realization of goal economic performance index local optimum of optimizing and then the economic performance index optimum of realization long-time running.

Each step in the rolling optimization is optimized employing SQP optimized Algorithm:

The SQP optimized Algorithm is described:

For the general nonlinearity optimization problem:

min f(x)

s.t.g _i(x)≥0，i＝{1，2，…，m ₁} (14)

h _j(x)＝0，j＝{m ₁+1，…m}

In the formula: $g_i\left(x\right)={(g_1\left(x\right),\&CenterDot;\&CenterDot;\&CenterDot;,g_{m_1}\left(x\right))}^T,$ $h_j\left(x\right)={(h_{{\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="A2009100562940022H1.png"\; state="\{\{state\}\}">m</figure-callout>}}_1+1}\left(x\right),\&CenterDot;\&CenterDot;\&CenterDot;,h_m\left(x\right))}^T$

(1) subproblem of structure QP

The QP subproblem can be described as:

$\min {\left[q\left(d\right)\right]}^{\left(k\right)}=\frac{1}{2}{d}^T{W}^{\left(k\right)}d+\&dtri;f{\left(x^{\left(k\right)}\right)}^{T}d$

$s.t.\&dtri;g_i{\left(x^{\left(k\right)}\right)}^{T}d+g_i\left(x\left(k\right)\right)\&GreaterEqual;0---\left(15\right)$

$\&dtri;h_E{\left(x^{\left(k\right)}\right)}^{T}d+h_E\left(x\left(k\right)\right)=0$

In the formula, W ^(k)Be extra large gloomy matrix, $W\left(k\right)=W(x^{\left(k\right)},{\mathrm{\&lambda;}}^{\left(k\right)})={\&dtri;}_x^{2}L(x^{\left(k\right)},{\mathrm{\&lambda;}}^{\left(k\right)}),$ Available symmetric positive definite matrix B ^(k)The approximate replacement, here,

$L(x,\mathrm{\&lambda;})=f\left(x\right)-\underset{i=1}{\overset{\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="A2009100562940022H1.png"\; state="\{\{state\}\}">m</figure-callout>}1}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_i{g}_i\left(x\right)-\underset{j=\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="A2009100562940022H1.png"\; state="\{\{state\}\}">m</figure-callout>}1+1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j{h}_j\left(x\right)---\left(16\right)$

The K-K-T condition is

$\&dtri;\mathrm{xL}(x^{*},{\mathrm{\&lambda;}}^{*})=0;$

h _j(x) ^*＝0，j∈E； (17)

${\mathrm{\&lambda;}}_i^{*}\&GreaterEqual;0,g_i\left(x^{*}\right)\&GreaterEqual;0,{\mathrm{\&lambda;}}_i^{*}{g}_i\left(x^{*}\right)=0,i\&Element;I.$

Can separate (d ^(k), λ ^(k+1));

(2) structure benefit function and descent direction

The definition benefit function:

$\mathrm{\&Phi;}1(x,\mathrm{\&mu;})=f\left(x\right)+\mathrm{\&mu;}[{\left|\right|\min \{g_i\left(x\right),0\}\left|\right|}_1+{\left|\right|h_E\left(x\right)\left|\right|}_1]=f\left(x\right)+\mathrm{\&mu;}\left(\underset{i\&Element;I}{\mathrm{\&Sigma;}}\min \right\{g_i\left(x\right),0\}+\underset{j\&Element;E}{\mathrm{\&Sigma;}}|h_j\left(x\right)|---\left(18\right)$

In the formula, μ is a penalty factor;

(3) step-length normal root certificate really:

Φ1(x ^(k)+α _kd ^(k)，μ ^(k))≤Φ ₁( ^x(k)，μ ^(k))+βα _kD(Φ ₁(x ^(k)，μ ^(k))；d ^(k)) (19)

Determine step-length α _k, β ∈ (0,1) wherein.

Concrete optimization step is:

1) determine objective function and constraint condition, objective function f (x) chooses formula (12) herein, constraint

Condition g _i(x) 〉=0, h _j(x)=0 as follows respectively with formula (13) corresponding relation:

$\left\{\begin{array}{c}{g}_1\left(x\right)=\mathrm{per}-\mathrm{perlolm}\&GreaterEqual;0\\ g_2\left(x\right)=\mathrm{perhilm}-\mathrm{per}\&GreaterEqual;0\end{array}\right.---\left(20\right)$

$\left\{\begin{array}{c}{h}_3\left(x\right)=\hat{E}-f_e(\mathrm{density},\mathrm{sp},\mathrm{nir},\mathrm{aro},\mathrm{per},\mathrm{ff},\mathrm{sdr},\mathrm{cop})=0\\ h_4\left(x\right)=\hat{P}-f_p(\mathrm{density},\mathrm{sp},\mathrm{nir},\mathrm{aro},\mathrm{per},\mathrm{ff},\mathrm{sdr},\mathrm{cop})=0\\ h_5\left(x\right)=\mathrm{per}-\hat{P}/\hat{E}=0\end{array}\right.---\left(21\right)$

With current cracking severity index promptly the third second ratio as initial value, choose SQP parameter μ with current vapour hydrocarbon ratio as initial value, δ, ξ＞0 (getting μ=0.05, δ=1, ξ=0.1), initial unit matrix B ₀, make k=0;

2) in restriction range, separate SQP subproblem (15), obtain separating d ^(k), λ ^(k+1), promptly optimum third second compares per _OptCompare sdr with the vapour hydrocarbon _Opt

7. the optimum cracking severity index that aforementioned calculation is obtained is that third second compares per _OptCarry out amplitude limiting processing.If cracking severity index third second that the field technician determines is [perlolm, perhilm] than restriction range, obtain attainable optimum cracking severity index third second ratio through amplitude limiting processing

${\hat{S}}_{\mathrm{opt}}=\left\{\begin{array}{c}\mathrm{perlolm},\mathrm{if}{\mathrm{per}}_{\mathrm{opt}}<\mathrm{perlolm}\\ {\mathrm{per}}_{\mathrm{opt}},\mathrm{ifperlolm}<{\mathrm{per}}_{\mathrm{opt}}<\mathrm{perhilm}\\ \mathrm{perhilm},\mathrm{if}{\mathrm{per}}_{\mathrm{opt}}>\mathrm{perhilm}\end{array}\right.---\left(22\right)$

8. consider the tracking performance of cracking severity depth controller, the variable quantity of the cracking severity setting value of this step output be provided be limited to a that third second of establishing current cracking severity controller is S than setting value _Now, then this step is optimized the setting value S that passes to controller down _New,

$S_{\mathrm{new}}=\left\{\begin{array}{cc}{S}_{\mathrm{now}}+a& \mathrm{if}({\hat{S}}_{\mathrm{opt}}-S_{\mathrm{now}})>a\\ s_{\mathrm{now}}-a& \mathrm{if}(S_{\mathrm{now}}-{\hat{S}}_{\mathrm{opt}})>a\\ {\hat{S}}_{\mathrm{opt}}& \mathrm{if}|S_{\mathrm{now}}-{\hat{S}}_{\mathrm{opt}}|<a\end{array}\right.---\left(23\right)$

Wherein the setting recommended range of the variation upper limit a of cracking severity setting value is 0.001～0.005.

9. if optimization cycle arrives, carry out next step computation optimization, promptly repeat step 4 to step 8, optimization cycle requires to determine that recommended range is 30 minutes～2 hours according to on-the-spot real-time.

Above-mentioned steps 1 belongs to early development stage of pyrolysis furnace yield model to step 3, and step 4 and step 9 belong to cracking severity on-line optimization process; By step 4 and step 9 as can be seen, on-line optimization is that the time span rolling is implemented with the optimization cycle, in each sampling instant, optimizes performance index and only relates to from this following limited constantly time, and to next sampling instant, this optimization period passes forward simultaneously.Though the optimization performance index form in each optimization cycle is identical, its time meaning difference, the first, the price factor in the objective function can be adjusted in real time according to market and change; Second, for this optimization problem, optimization variable is COT and vapour hydrocarbon ratio, in each optimization cycle, other model parameters comprise material density density, raw material straight-chain paraffin total amount sp, the ratio nir of raw material n-alkane isoparaffin content, material flow ff, cop is a boiler tube top hole pressure etc., because the real time data of enchashment field, and constantly change, promptly in each optimization cycle, though the optimization aim form is identical, owing to model parameter constantly changes, then under the prerequisite of objective function maximum, optimisation strategy is in time adjusted in variation according to actual production situation, promptly adjusts COT and vapour hydrocarbon ratio.

Only for the preferred embodiment of invention, be not to be used for limiting practical range of the present invention in sum.Be that all equivalences of doing according to the content of the present patent application claim change and modification, all should be technology category of the present invention.

Claims (10)

1. method for optimizing cracking depth of industrial ethane cracking furnace on line is characterized in that said method comprising the steps of:

1. gather the historical floor data and the corresponding oil product attribute data of industrial ethylene pyrolysis furnace system;

2. choose the oil product attribute as the cluster attribute variable, carry out the oil product cluster: utilize the fuzzy C-means clustering algorithm that the oil product data are classified as the similar N class of crack characteristic;

3. according to the oil product cluster result of step in 2., utilize mechanism model to generate the data set of " raw material attribute-operating conditions-product yield " corresponding with N class oil product, N the neural network model that structure has the nonlinear Identification ability, wherein, mechanism model is the Simultaneous Equations that comprises cracking reaction kinetics equation, mass-conservation equation group, momentum conservation equation, energy conservation equation;

4. gather current working data and corresponding oil product attribute data,, calculate in conjunction with current oil product property value

This oil product is to the corresponding oil product distances of clustering centers of each submodel, the chosen distance minimum be current model: $d\left(i\right)=\sqrt{\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}{|x_j-c_i^j|}^2},$ Selecting the minimum submodel of d (i) is current model, and wherein, the current measurement oil product of d (i) expression is to the corresponding oil product distances of clustering centers of each submodel, and the cluster centre of the corresponding oil product of each submodel is (c _i ¹, c _i ²... c _i ^j), current oil product property value is (x ₁, x ₂..., x _j), i=1 ... N, j=1 ... M, N are the oil product cluster classification number of step described in 2., and M is the cluster attribute variable number of step described in 2.;

5. for overcoming the uncertainty that model mismatch and external disturbance can bring to system usually, adopt keeping and on the forecast model basis of invariable error in future is made a prediction and carry out on-line correction: y _p(k+1)=y _m(k+1)+he (k), wherein, y _p(k+1) for next third second constantly of the k+1 after proofreading and correct than model predication value, y _m(k+1) be next third second constantly of k+1 than model prediction computation value, h is the error correction coefficient, e (k)=y (k)-y _mThird second of etching system is than actual value y (k) and model predication value y when (k) being k _m(k) error between;

6. according to the actual conditions of pyrolysis furnace, determine Optimizing operation variable and constraint condition, be target to the maximum with ethylene, propylene mass yield sum, predict that with the pyrolysis product yield behind the on-line correction neural network model is as the on-line optimization model, utilize the SQP method in conjunction with near the expert system rolling optimization cracking severity desired value of current C OT working point correspondence, obtain cracking severity index optimization net result.

2. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1 is characterized in that: the floor data of described step in 1. comprises the feed rate of pyrolysis furnace, steam flow, outlet temperature COT, top hole pressure COP, sidewall fuel gas flow, bottom fuel gas flow; The oil product attribute comprises the quality percentage composition of n-alkane, isoparaffin, alkene, naphthenic hydrocarbon, aromatic hydrocarbons, methane, ethene, propylene content number percent in the pyrolysis gas.

3. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1 is characterized in that: described historical data acquisition range is historical floor data and the corresponding oil product attribute data in the period of 3 months to 2.

4. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1 is characterized in that: described step 2. in, choose likening to of straight-chain paraffin total amount, n-alkane and isoparaffin and be the cluster attribute variable.

5. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1, it is characterized in that: 2. described step is by adopting the random initializtion cluster centre and repeatedly move the fuzzy C-means clustering algorithm, selecting the minimum corresponding result of cost function to realize the oil product data qualification.

6. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1 is characterized in that: the operating conditions that the data set of " the raw material attribute-operating conditions-product yield " of described step in 3. produces is 100%～120% of nominal situation scope basis.

7. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1, it is characterized in that: the model of described step in 3. is the neural network regression model, wherein, each sub neural network model input variable all is chosen for: material density, raw material straight-chain paraffin content sum, the ratio of raw material n-alkane isoparaffin content, feed rate, vapour hydrocarbon ratio, coil outlet temperature COT, boiler tube top hole pressure COP, the model output variable is chosen for: yield of ethene and propene yield.

8. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1 is characterized in that: the error correction coefficient value scope of described step in 4. is [0.5,1.0].

9. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1 is characterized in that: the 6. middle cracking severity index optimization net result of described step, passed through amplitude limiting processing.

10. method for optimizing cracking depth of industrial ethane cracking furnace on line according to claim 1, it is characterized in that: the rolling optimization of described step in 5., promptly in each sampling instant, optimizing performance index only related to from this following limited constantly time, and to next sampling instant, this optimization period passes forward simultaneously, constantly the optimization performance index in this moment is optimized at each.

Images