CN101882239A - Ethylene cracking severity modeling method based on expert knowledge and neutral network - Google Patents
Ethylene cracking severity modeling method based on expert knowledge and neutral network Download PDFInfo
- Publication number
- CN101882239A CN101882239A CN 201010188121 CN201010188121A CN101882239A CN 101882239 A CN101882239 A CN 101882239A CN 201010188121 CN201010188121 CN 201010188121 CN 201010188121 A CN201010188121 A CN 201010188121A CN 101882239 A CN101882239 A CN 101882239A
- Authority
- CN
- China
- Prior art keywords
- neural network
- cracking
- model
- variable
- expertise
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention relates to a modeling method for soft sensing the cracking severity of an ethylene cracking furnace based on expert knowledge and a neutral network. A model established by using the method has high prediction precision, high reliability and high extrapolation performance. The model can reflect the actual operation characteristics of the cracking furnace accurately in real time so as to instruct an operator to adjust the operation variables of the cracking process in time, and thus, the economical benefits are increased. In the invention, the expert knowledge about the ethylene cracking process is added to the network training process of ethylene cracking severity neutral network modeling, and a training sample set is formed by acquiring and preprocessing onsite production data. Meanwhile, the neutral network is optimized by using an intelligent evolutionary algorithm, and the neutral network soft sensing model of ethylene cracking severity is established.
Description
Technical field
The invention belongs to chemical engineering and information science crossing domain, relate to a kind of ethylene cracking severity soft-measuring modeling method based on expertise and neural network.
Background technology
Ethene is the important foundation raw material of chemical industry, and its output becomes the outstanding feature of weighing a country petrochemical complex development.Along with the continuous expansion energy of ethylene unit and going deep into of the integrated work of oil-refining chemical, the ethylene raw resource is nervous and the diversified problem of raw material is outstanding day by day, in cracker maximization, diversified while of raw material, the feed flexibility, raising ethylene selectivity and yield, minimizing energy consumption and the material consumption that improve cracker just seem particularly important.
Pyrolysis furnace is the nucleus equipment of ethylene unit, and can pyrolysis furnace efficiently move, and is directly connected to the production cost and the whole economic efficiency of ethylene unit.Because raw material sources and the often change of production load, only the operating conditions of pyrolysis furnace is adjusted with knowhow, often cause triolefin (referring to ethene, propylene and butadiene) total recovery low, ethane cracking furnace moves departing under the state of optimal operating condition usually, and material consumption height, energy consumption are big.Therefore,, the cracking operation condition is optimized control, can guarantees the stable of product quality, improve the triolefin yield, improve the production competitive power of enterprise, and can produce huge economic benefit according to production status.
Cracking severity is the key control variable of ethane cracking furnace, and cracking severity has reflected yield of ethene and triolefin total recovery to a certain extent.The too low yield of ethene that influences of ethylene cracking severity, the too high speed of then accelerating tube coking, and then the cycle of operation of shortening pyrolysis furnace influence economic benefit.Feedstock oil is Millisecond in the residence time of boiler tube in the ethane cracking furnace, and the timely adjustment to process just seems particularly important like this.At the measurement of cracking severity, though the general industry scene all is equipped with analysis meter, 20-30 minute hysteresis is arranged all, and safeguard complicated, with high costs.Lag behind for a long time simultaneously the residence time extremely short with respect to cracking stock, if process is adjusted according to the result of analysis meter, the very difficult pyrolysis furnace that guarantees is worked under optimal operation conditions again.For this reason, propose to have a kind of preferably modeling method that merges the soft measurement of pyrolysis furnace cracking severity of expertise of rapid reaction, extrapolability.
Summary of the invention
The purpose of this invention is to provide a kind of ethylene cracking severity neural network modeling approach that merges expertise.Choose dilution steam generation flow, feedstock oil feed rate, feedstock oil density, boiler tube outlet medial temperature, waste heat boiler outlet medial temperature and radiation section medial temperature as the model input variable, choose the dependent variable of the pyrolysis furnace cracking reaction degree of depth, adopt nerual network technique to merge the correlation model that expertise is set up input variable and output variable as model; The model of setting up has higher forecast precision, good model extrapolability is arranged again simultaneously.But the on-line correction of network output binding analysis instrument provides cracking severity in real time, realizes the soft measurement to ethylene cracking severity.And then realization brings economic benefit to the timely adjustment of cracking process operation variable.Its principle of work is that the information that some expertises or Analysis on Mechanism about the ethane cracking furnace operating characteristic obtain is incorporated in the training of neural network, utilize these prioris to instruct the modeling process of neural network, use intelligent evolution algorithm that the neural network error function is carried out the minimal value optimizing simultaneously, obtain the neural network model parameter.
It is the Sigmoid function that the soft measurement neural network model of cracking severity adopts activation function, and three layers of feedforward neural network are set up, 6 nodes of input layer, 5~10 nodes of hidden layer, 1 node of output layer.
The input variable of cracking severity neural network model:
(1) dilution steam generation flow (x
1, kg/h)
(2) cracking stock oil feed rate (x
2, kg/h)
(3) feedstock oil density (x
3, kg/m
3)
(4) boiler tube outlet medial temperature (x
4, ℃)
(5) waste heat boiler outlet medial temperature (x
5, ℃)
(6) radiation section medial temperature (x
6, ℃)
The dependent variable of model, the i.e. output variable of neural network model: cracking severity.The input variable of neural network model utilizes following formula (1) to carry out normalized:
In the formula (1), x
iBe the measurement data of i operating conditions, sx
iRepresent i operating conditions input value as neural network after normalized,
The lower limit and the upper limit of i operating conditions variation that expression collects, a and b represent the lower limit and the upper limit of input variable after the normalized.
The output variable of network model utilizes following formula (2) to carry out normalized:
In the formula (2), y is the actual measured value of output variable, y
Min, y
MaxBe the lower limit and the upper limit of the neural network output variable actual measured value set, the later value of normalization is sy, and c and d represent the lower limit and the upper limit of the output valve of neural network model after the normalized.
Collect n and organize representational test data, wherein every group of data comprise [x
1, x
2..., x
6, y], after normalization [sx
1, sx
2..., sx
6, sy], form training sample.Cracking reaction depth neural network model is trained, with [sx
1, sx
2..., sx
6] as the input of network, be desired value with sy, training network.
The training of neural network is a process that error function is minimized, traditional training method with square error as the error criterion function of weighing the network capability of fitting:
Wherein, N is the number of training sample, sy
iBe the value after i the sample normalization, y '
iIt is the output valve of i sample neural network.Merge the neural metwork training process of expertise, be exactly to add in the training stage of neural network the sensitivity analysis of the crucial input variable of model is judged, in the training process of model, each iteration is carried out sensitivity analysis to the main input variable (this patent refers to pyrolysis furnace feed rate, steam flow, coil outlet temperature) of model.The residence time of pyrolysis furnace feed rate reflection reactant, steam flow and inlet amount reflection vapour hydrocarbon ratio, coil outlet temperature reflection cracking reaction degree.If the variation tendency of the corresponding output variable of the ethane cracking furnace that model output that sensitivity analysis obtains and expertise or Analysis on Mechanism provide is consistent, then the error function to neural network does not add punishment.Otherwise, this error function is imposed suitable punishment.The training flow process of neural network as shown in Figure 2.The error formula of introducing the neural metwork training of sensitivity analysis can be written as:
In the formula (4), f
jBe the sensitivity analysis zone bit of j key variables, if the sensitivity analysis of j key variables is consistent with the mechanism result, f then
j=0, otherwise, f
j=1.J key variables are imposed suitable punishment.p
jIt is the penalty term of j crucial input variable.The variable number that will carry out sensitivity analysis of M for determining according to expertise, here, we only choose the boiler tube outlet medial temperature of steam flow, feed rate and the pyrolysis furnace of pyrolysis furnace and carry out sensitivity analysis, so M=3.
Neural network training adopts intelligent evolution algorithm (as genetic algorithm, particle swarm optimization algorithm, differential evolution algorithm etc.), when reaching accuracy requirement, stops training, and the weights of the neural network model that obtains and threshold value are determined the model of pyrolysis furnace cracking severity.
Description of drawings
Fig. 1 is a cracking severity neural network model block diagram.Fig. 2 adopts the evolution algorithm training to merge the pyrolysis furnace cracking severity soft-sensing model process flow diagram based on neural network of expertise.
Embodiment
Below, the invention will be further described by embodiment, and it will help to understand the present invention, but not limit content of the present invention.
Embodiment
Adopt neural network structure as shown in Figure 1, determine that at first the variable that influences ethylene cracking severity has: dilution steam generation flow (x
1, kg/h), pyrolysis furnace feed loading (x
2, kg/h), feedstock oil density (x
3, kg/m
3), boiler tube outlet medial temperature (x
4, ℃), waste heat boiler outlet medial temperature (x
5, ℃), radiation section medial temperature (x
6, ℃).This example is to crucial input variable dilution steam generation flow x
1, pyrolysis furnace feed loading x
2With boiler tube outlet medial temperature x
4Do sensitivity analysis.So determine that according to above variable the input layer number of neural network is 6, the hidden layer node number is made as 8, output layer node number is 1.The hidden layer activation function adopts the tansig function, and the output layer activation function adopts the logsig function.
According to input, the output variable of top definite neural network, to gather from totally 2000 groups of the data of industry spot, these data have covered different operating conditions, can comparatively comprehensively react the operation characteristic of pyrolysis furnace.According to (1) formula, above-mentioned each independent variable is carried out normalized, all independents variable all normalize within [1,1] scope, so (1) formula is got a=-1, b=1.Input variable x
1Bound be made as [14192,14362], x
2Bound be made as [20992,25153], x
3Bound be made as [619.54,657.46], x
4Bound be made as [832.46,852.92], x
5Bound be made as [410.08,419.84], x
6Bound be made as [591.25,629.99], normalization is calculated as follows:
Output variable ethylene cracking severity y (C
3/ C
2) bound be made as [0.5392,0.6207], the normalization scope of output variable is [0.1,0.9], the normalization formula is as follows:
To the sensitivity analysis conclusion of dilution steam generation flow, pyrolysis furnace feed loading and boiler tube outlet medial temperature, this example with its change to cracking severity influence rule as expertise, introduce in the middle of the training process of neural network.The error formula of neural metwork training is written as:
Wherein, y '
iBe the neural network output valve of i group sample, sy
iBe the value of i group sample actual measured value after normalization, f
1, f
2, f
3Be respectively key variables steam flow x
1; Pyrolysis furnace feed loading x
2With pyrolysis furnace outlet temperature x
4The punishment zone bit.p
1, p
2And p
3Be taken as 10,10 and 20 respectively, the p here
iGet different values according to variable importance.
The neural metwork training process that merges expertise as shown in Figure 2, the concrete step of the training process of neural network is as follows:
Step1: the weights of initialization neural network and threshold value.Wherein input layer to the weights of hidden layer is
(i=1,2 ..., 6; J=1,2 ..., 8), hidden layer to the output layer weights is
(i=1, j=1,2 ..., 8),
Step2: output and actual comparison with neural network obtain error term E
e
Step3: utilize steam variable, pyrolysis furnace feed loading and three key variables of pyrolysis furnace outlet temperature to do sensitivity analysis to the current model that obtains, carry out monotonicity and judge.As judging when the steam flow increase, whether the output of model is dull decline., dullness puts punishment zone bit f if descending
1=0, otherwise, f made
1=1.Carry out the sensitivity analysis of second and the 3rd variable equally, obtain punishing the value of zone bit.Thus, obtain the penalty term E of objective function
p
Step4: calculate total error function E=E
e+ E
p, utilize intelligent evolution algorithm, adjust weights and threshold value error function is optimized.
Step5: the end condition that judges whether to meet training.Carry out Step6 if satisfy, otherwise, forward Step2 to.
Step6: the weights of output nerve network and threshold value, determine final cracking severity soft-sensing model.
Claims (10)
1. ethylene cracking severity modeling method based on expertise and neural network, it is characterized in that: described flexible measurement method comprises the steps:
(1) chooses the input variable of the performance variable of ethane cracking furnace as the ethylene cracking severity neural network model;
(2) choose the output variable of ethylene cracking severity (mass ratio of propylene and ethene) as model;
(3) by gathering the actual production data of ethane cracking furnace, after removing gross error, form training sample set, collect the expertise and the existing mechanism derivation conclusion of cracking of ethylene reaction, form expert knowledge library, utilize intelligent evolution algorithm neural network to be trained in conjunction with expertise, set up the neural network model of ethylene cracking severity, according to the cracking severity decision operation operating mode of model output, timely adjustment process performance variable, the triolefin yield of assurance ethane cracking furnace.
2. modeling method according to claim 1 is characterized in that, described step (1) and (2) are to the neural net model establishing technology of cracking of ethylene course of reaction, adopt three layers of feedforward neural network, the input layer number is 6, the corresponding dilution steam generation flow of difference, feedstock oil feed rate, feedstock oil density, boiler tube outlet medial temperature, waste heat boiler outlet medial temperature and radiation section medial temperature, the hidden layer node number is 5~10, output layer node number is 1, corresponding cracking severity output valve.
3. modeling method according to claim 1, it is characterized in that, to be incorporated into the training process of neural network about the expertise of cracking of ethylene process in the described step (3), the expertise of choosing is mainly the crucial input variable of cracking process is carried out sensitivity analysis to output variable change information.
4. modeling method according to claim 2 is characterized in that the input of described neural network, output variable are all through normalized.
5. modeling method according to claim 4 is characterized in that, the input variable of described neural network model is carried out normalized according to following formula:
In the formula, x
iBe the measurement data of i operating conditions, sx
iRepresent i operating conditions input value as neural network after normalized,
The lower limit and the upper limit of i the operating conditions variation that expression is gathered, a and b represent the lower limit and the upper limit of input variable after the normalized;
The dependent variable of described neural network model utilizes following formula to carry out normalized:
In the formula, y is the actual measured value of output variable, y
Min, y
MaxBe the lower limit and the upper limit of the output variable actual measured value set, the later value of normalization is sy, and c and d represent the lower limit and the upper limit of the output valve of neural network model after the normalized.
6. modeling method according to claim 5 is characterized in that, with the later data [sx of normalization
1, sx
2..., sx
8, sy] and as training sample, in the iteration searching process, each iteration is carried out the sensitivity analysis of crucial input variable to the gained model, and whether the comparative analysis result is consistent with expertise or mechanistic information, and inconsistent model error function is applied penalty term.
7. modeling technique according to claim 6, it is characterized in that, the sensitivity analysis and the expertise of gained model are compared judgement, judged result writes the neural network error function with the form of penalty term, and then expertise participates in the iteration each time of neural metwork training with training sample.
8. modeling technique according to claim 7, it is characterized in that, adopt intelligent evolution algorithm that feedforward neural network is trained, the intelligence evolution algorithm can adopt genetic algorithm, particle swarm optimization algorithm, differential evolution algorithm etc., during training, the neural network error function be two parts add and, wherein, a part is the error sum of squares E of the output y ' of the actual value sy of sample and neural network
e, another part is for being judged the penalty term E that is introduced by expertise
p, when error function reaches minimum value, obtain the correlation parameter such as weights, threshold value of neural network.
9. modeling technique according to claim 1 is characterized in that, the model that obtains need carry out anti-normalized could get to the end cracking severity output valve, and anti-normalization is calculated by following formula:
In the formula,
Be the output valve of the soft measurement of cracking severity, y
Min, y
MaxThe lower limit and the upper limit for the neural network output variable actual measured value set, identical with in the claim 5, y ' is the calculated value of neural network, and c and d represent the lower limit and the upper limit of the output valve normalized of neural network model, identical with in the claim 5.
10. modeling technique according to claim 1 is characterized in that, when prediction, with corresponding input variable input neural network, being exported accordingly is ethylene cracking severity, and operating personnel can judge according to the numerical value of cracking severity whether operating mode is normal.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN 201010188121 CN101882239A (en) | 2010-05-28 | 2010-05-28 | Ethylene cracking severity modeling method based on expert knowledge and neutral network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN 201010188121 CN101882239A (en) | 2010-05-28 | 2010-05-28 | Ethylene cracking severity modeling method based on expert knowledge and neutral network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN101882239A true CN101882239A (en) | 2010-11-10 |
Family
ID=43054251
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN 201010188121 Pending CN101882239A (en) | 2010-05-28 | 2010-05-28 | Ethylene cracking severity modeling method based on expert knowledge and neutral network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN101882239A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103087758A (en) * | 2011-10-28 | 2013-05-08 | 中国石油化工股份有限公司 | Naphtha industrial cracking furnace value maximization model construction method |
CN103093069A (en) * | 2011-10-28 | 2013-05-08 | 中国石油化工股份有限公司 | Construction method of value-maximization model of industrial cracking furnace |
CN103087759A (en) * | 2011-10-28 | 2013-05-08 | 中国石油化工股份有限公司 | Industrial cracking furnace cracking product value maximization model construction method |
CN103207949A (en) * | 2013-04-10 | 2013-07-17 | 华东理工大学 | Predicating method for TLE (transfer line exchanger) outlet temperatures and operation cycles of ethylene cracking furnaces |
CN103605325A (en) * | 2013-09-16 | 2014-02-26 | 华东理工大学 | Complete period dynamic optimization method for industrial ethylene cracking furnace and based on surrogate model |
CN104715142A (en) * | 2015-02-06 | 2015-06-17 | 东南大学 | NOx emission dynamic soft-sensing method for power station boiler |
CN104915522A (en) * | 2015-07-01 | 2015-09-16 | 华东理工大学 | Mixed modeling method and system based on combination of process priors and data-driven model |
CN108681656A (en) * | 2018-04-27 | 2018-10-19 | 上海卓然工程技术股份有限公司 | A kind of process analysis method based on ethane cracking furnace operation data |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101017535A (en) * | 2007-03-08 | 2007-08-15 | 上海交通大学 | Implementation method of retina encoder using space time filter |
-
2010
- 2010-05-28 CN CN 201010188121 patent/CN101882239A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101017535A (en) * | 2007-03-08 | 2007-08-15 | 上海交通大学 | Implementation method of retina encoder using space time filter |
Non-Patent Citations (1)
Title |
---|
《计算机与应用化学》 20090828 尚田丰等 基于GA-RBF网络的乙烯裂解炉在线操作优化 1003-1006 1-10 第26卷, 第8期 2 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103087758B (en) * | 2011-10-28 | 2014-12-31 | 中国石油化工股份有限公司 | Naphtha industrial cracking furnace value maximization model construction method |
CN103093069A (en) * | 2011-10-28 | 2013-05-08 | 中国石油化工股份有限公司 | Construction method of value-maximization model of industrial cracking furnace |
CN103087759A (en) * | 2011-10-28 | 2013-05-08 | 中国石油化工股份有限公司 | Industrial cracking furnace cracking product value maximization model construction method |
CN103093069B (en) * | 2011-10-28 | 2016-09-21 | 中国石油化工股份有限公司 | A kind of construction method of industrial pyrolysis furnace value maximization model |
CN103087758A (en) * | 2011-10-28 | 2013-05-08 | 中国石油化工股份有限公司 | Naphtha industrial cracking furnace value maximization model construction method |
CN103207949B (en) * | 2013-04-10 | 2016-09-07 | 华东理工大学 | Ethane cracking furnace TLE outlet temperature and cycle of operation Forecasting Methodology |
CN103207949A (en) * | 2013-04-10 | 2013-07-17 | 华东理工大学 | Predicating method for TLE (transfer line exchanger) outlet temperatures and operation cycles of ethylene cracking furnaces |
CN103605325B (en) * | 2013-09-16 | 2015-09-09 | 华东理工大学 | Based on the industrial ethylene pyrolysis furnace complete period dynamic optimization method of agent model |
CN103605325A (en) * | 2013-09-16 | 2014-02-26 | 华东理工大学 | Complete period dynamic optimization method for industrial ethylene cracking furnace and based on surrogate model |
CN104715142A (en) * | 2015-02-06 | 2015-06-17 | 东南大学 | NOx emission dynamic soft-sensing method for power station boiler |
CN104715142B (en) * | 2015-02-06 | 2018-02-02 | 东南大学 | A kind of station boiler NOxDischarge dynamic soft-measuring method |
CN104915522A (en) * | 2015-07-01 | 2015-09-16 | 华东理工大学 | Mixed modeling method and system based on combination of process priors and data-driven model |
CN104915522B (en) * | 2015-07-01 | 2019-06-25 | 华东理工大学 | The hybrid modeling method and system of cohesive process priori and data-driven model |
CN108681656A (en) * | 2018-04-27 | 2018-10-19 | 上海卓然工程技术股份有限公司 | A kind of process analysis method based on ethane cracking furnace operation data |
CN108681656B (en) * | 2018-04-27 | 2023-09-01 | 上海卓然工程技术股份有限公司 | Process analysis method based on ethylene cracking furnace operation data |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN101882239A (en) | Ethylene cracking severity modeling method based on expert knowledge and neutral network | |
CN101620414B (en) | Method for optimizing cracking depth of industrial ethane cracking furnace on line | |
Liang et al. | Dynamic evaluation of drilling leakage risk based on fuzzy theory and PSO-SVR algorithm | |
CN101414158A (en) | Method for optimizing cracking reaction operating condition of ethylene cracking furnace | |
CN104789256B (en) | A kind of yield real-time predicting method of catalytic cracking unit | |
CN103524284B (en) | Forecasting and optimizing method for ethylene cracking material configuration | |
CN102053595A (en) | Method for controlling cracking depth of cracking furnace in ethylene device | |
CN102213949B (en) | Method for optimizing value of ethylene device | |
CN104484714B (en) | A kind of real-time predicting method of catalytic reforming unit yield | |
CN109814513A (en) | A kind of catalytic cracking unit optimization method based on data model | |
CN102626557B (en) | Molecular distillation process parameter optimizing method based on GA-BP (Genetic Algorithm-Back Propagation) algorithm | |
CN106202946A (en) | Clinker free calcium levels Forecasting Methodology based on degree of depth belief network model | |
CN102184287B (en) | Modelling method for combustion optimization of waste plastics oil refining | |
CN101349893A (en) | Forecast control device of adaptive model | |
CN100334047C (en) | Intelligent method for controlling cracking severity of cracking furnace in ethylene equipment | |
CN104765346A (en) | Full-process modeling method for oil refining process | |
CN104463327A (en) | Method for predicting catalytic cracking coke yield | |
CN102175345A (en) | Soft measurement method for fire box temperature of multi-nozzle opposed coal water slurry gasification furnace | |
CN104361153A (en) | Method for predicting coking amount of heavy oil catalytic cracking settler | |
CN110376987A (en) | The oil refining process method for on-line optimization and device detected based on case library and stable state | |
CN104765347A (en) | Yield real-time prediction method in residual oil delayed coking process | |
CN105868867A (en) | Method and system for optimized operation of heating boiler cluster | |
CN111639820A (en) | Energy consumption parameter optimization method and system for cement kiln production | |
CN103323484A (en) | Method for predicting crystallization state of sugarcane sugar boiling process | |
CN101727609B (en) | Pyrolyzate yield forecasting method based on support vector machine |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20101110 |