CN110322933A - A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism - Google Patents
A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism Download PDFInfo
- Publication number
- CN110322933A CN110322933A CN201910534700.1A CN201910534700A CN110322933A CN 110322933 A CN110322933 A CN 110322933A CN 201910534700 A CN201910534700 A CN 201910534700A CN 110322933 A CN110322933 A CN 110322933A
- Authority
- CN
- China
- Prior art keywords
- sample
- training
- model
- new
- data
- 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
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/30—Prediction of properties of chemical compounds, compositions or mixtures
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/70—Machine learning, data mining or chemometrics
Landscapes
- Engineering & Computer Science (AREA)
- Chemical & Material Sciences (AREA)
- Theoretical Computer Science (AREA)
- Computing Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Crystallography & Structural Chemistry (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism, raw sample data collection is divided into training sample set and test sample collection first, mechanism model is established using training set, multiple new training sample subsets are established with random method for resampling again, each subset of training establishes the base learner set of Elman neural network;Then clustering is carried out, choosing the preferable individual of performance in every cluster, alternatively property integrates subset;The hybrid modeling method in parallel of the dynamic compensation mechanism is finally used for test sample collection, whether needs to update mechanism model training sample and corresponding parameter according to similarity principle using error judgment, obtains corresponding polypropylene melt index predicted value.Precision of prediction of the present invention is high, highly reliable, according to obtained hard measurement as a result, it is possible to preferably track the melt index Nonlinear Dynamic variation tendency of propylene polymerization production process, provides effective technical support for polypropylene production process operation optimization and quality control.
Description
Technical field
The invention belongs to the research of polypropylene production process flexible measurement method and application fields, more particularly to one kind is based on dynamic
The polypropylene melt index hybrid modeling method of error compensation mechanism.
Background technique
With the increasingly complexity of industrial process object, to control product quality, regular industrial monitoring technology is difficult to reach again
Satisfactory result.The inspection of many offline samples not only time-consuming cost, but also control is easily caused to lag, product quality and life
Safety is produced all to be difficult to be protected.Therefore, soft-measuring technique is introduced chemical production process has important display meaning.
In chemical process and the cross-application of Other subjects, based on the mechanism model that chemical industry object is established, accuracy
By actual condition, the factors such as operation cost and and actual conditions there are certain deviations.Data-driven model by secret operation without
Priori knowledge is needed to can be obtained the mapping relations between input and output.Currently, data-driven model is in chemical industry city with keen competition
It is becoming increasingly popular in, reference literature 1:Kadlec P, Gabrys B, Strandt S.Data-driven soft
sensors in the process industry[J].Computers and Chemical Engineering,2009,33
(4): 795-814. multiple linear regression, artificial neural network, support vector machines etc. constantly expand with field.Reference literature 2:
Coking process energy consumption prediction model [J] the computer and applied chemistry of Wei Na, Li Li selectivity Artificial neural network ensemble, 2013,
30 (10): 1127-1130. establishes selective compositive neural network model using cross validation back-and-forth method, and successfully realizes burnt
Change process energy consumption prediction.Reference literature 3: Xiong Weili, Yao Le and its are fermenting at Xu Baoguo chaos least square method supporting vector machine
Application [J] Journal of Chemical Industry and Engineering in process model building, 2013,64 (12): it is minimum that 4585-4591. proposes a kind of chaos algorithm optimization
Two multiply support vector machines hard measurement modeling method, effectively demonstrate for penicillin fermentation process.
But since practical chemical system has the characteristics that high complexity, strong nonlinearity, serious coupling and time variation more,
The single data-driven model of tradition is difficult to leading variable Accurate Prediction in actual condition.The precision of model predictive process is not high
And stability is not strong, global prediction poor performance, to actual production process prediction, there are certain deviations.
Summary of the invention
In order to overcome existing polypropylene melt index soft-sensing model not can guarantee precision of prediction, mould in global scope
The deficiencies of type stability is not strong, the present invention provide a kind of precision of prediction is higher, model stability it is stronger based on dynamic error mend
Repay the polypropylene melt index hybrid modeling method of mechanism.
The technical solution adopted by the present invention to solve the technical problems is:
A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism, the method includes following
Step:
1) according to actual industrial production operating condition, auxiliary variable and leading variable is selected, Distributed Control System and experiment are passed through
Analytical database acquires production operation data, establishes raw sample data collection X;
2) abnormal data in the original sample collection X that step 1) obtains is rejected using Lai Yite criterion, is disappeared using filtering method
Except the random error in data, and data normalized is carried out, eliminates data dimension and difference in size brings model training
Influence;
3) according to setting ratio, the pretreated raw sample data collection X of step 2) is divided into training sample set XMWith
Test sample collection XK;
4) training sample X is utilizedMMiddle auxiliary variable and leading variable establish mechanism mould according to the mechanism knowledge of production process
Type, and model parameter is recognized with Least-squares minimization algorithm;
5) by training sample set XMAuxiliary variable and the obtained error of modelling by mechanism respectively as data-driven error
Compensation model is output and input, and forms new training sample S;
6) to training sample S, extraction yield p and resampling number c is determined, is repeated at random using Bootstrap algorithm
Sampling, establishes several training sample subsets { S1, S2, S3..., Sc};
7) training parameter of Elman neural network is set, is trained based on various kinds book the set pair analysis model, establishes several
Elman neural network submodel;
8) Fuzzy C-Means Clustering Algorithm is used, clusters number is determined, screens the best individual conduct of accuracy in every cluster
Selective ensemble submodel is for integrating;
9) setting allows estimated bias lower limit, in new polypropylene production status, utilizes the good Elman mind of learning training
Evaluated error through network error compensation model prediction mechanism model and real system process, if error meets setting value, directly
Connect the output valve of output-parallel mixed model;
10) similarity for otherwise needing to calculate current production status sample data and historical data again, judges whether to meet
Sample Similarity lower limit such as meets condition, then does not need to be added in training set;
If 11) be unsatisfactory for the condition of Sample Similarity lower limit, need to determine the ginseng of more new model using cross-validation method
Several new number of training, and step 4) is repeated, mechanism model relevant parameter is updated, realizes known operation condition
The Accurate Prediction of lower polypropylene melt index.
Further, the treatment process of the step 6) are as follows:
6.1) error between the predicted value for obtaining experimental analysis value and modelling by mechanism is as data-driven error compensation
The output vector x of modeli=[xi,1,xi,2,…,xi,n]T∈Rn;
6.2) original training sample collection X is chosenMMiddle auxiliary variable yi=[yi,1,yi,2,…,yi,m]T∈RmIt is driven as data
The input of dynamic error compensation model;
6.3) to new training sampleDetermine that extraction yield is 80%, resampling number 10 uses
Bootstrap algorithm carries out random sampling with replacement, establishes 10 training sample subset { S1, S2, S3..., S10}。
Further, the treatment process of the step 7) are as follows:
7.1) to the training sample subset { S provided1, S2, S3..., Sc, Elman neural network uses 3-tier architecture, chooses
Node in hidden layer, and the activation primitive of tansing type and s shape function as hidden layer is chosen respectively;
7.2) for Elman neural network,Respectively input layer to hidden layer, accept layer to hidden layer connection weight
Weight matrix, f () indicate the excitation function of output unit, then corresponding hidden layer output are as follows:
In formula,Respectively connection weight matrix of the input layer to hidden layer, undertaking layer to hidden layer, f ()
Indicate the excitation function of output unit.
The output for accepting layer indicates are as follows:
In formula, t indicates the moment,Respectively input layer to hidden layer, accept layer to hidden layer connection weight
Matrix, f () indicate the excitation function of output unit.
7.3) the nonlinear state equation of output layer indicates are as follows:
In formula,Indicate connection weight matrix of the hidden layer to output layer, the excitation function of the implicit layer unit of g () expression.
Further, the treatment process of the step 8) are as follows:
8.1) individual of training sample subset is predicted to obtain using the Elman neural network submodel after training pre-
Survey output matrix H;
8.2) using Fuzzy C-Means Clustering Algorithm to prediction output matrix H clustering, objective function is indicated are as follows:
In formula, mn is clusters number;N is sample number;B is fuzzy clustering index, takes b >=1;υj(j=1,2 ..., m) be
The center of j-th of cluster;uj(xi) indicate i-th of sample for jth class degree of membership, and meet:
8.3) the cluster centre matrix V=[υ obtainedj] and subordinated-degree matrix U=[uj(xi)] iterative formula be respectively as follows:
8.4) it by obtained subordinated-degree matrix U, obtains each Elman neural network and angle value is subordinate to every cluster;
8.5) each Elman neural network submodel is belonged in the maximum cluster of degree of membership, and according to average relative error
Minimum principle chooses the higher Elman neural network submodel of the accuracy alternatively integrated subset of property from every cluster.
The treatment process of the step 10) are as follows:
10.1) current production status sample data x is calculatednewWith history training sample data XMMiddle auxiliary variable data yi=
[yi,1,yi,2,…,yi,m]T∈RmThe formula of distance and angle is respectively as follows:
In formula, l (xnew,xm) and θ (xnew,xm) respectively indicate the calculation formula of distance and angle;
10.2) according to distance and angle, the similarity of current production status sample data and historical data is obtained, is counted
Calculate formula are as follows:
Snew,i=β exp (- l (xnew,yi))+(1-β)cosθ(xnew,yi)
In formula, β indicates weight coefficient and 0≤β≤1;
10.3) if the similarity obtained meets given Sample Similarity lower limit, do not need to say that it is added to training set
In sample.
The treatment process of the step 11) are as follows:
11.1) if similarity obtained in step 10) is unsatisfactory for given Sample Similarity lower limit, need to update training
Sample number;
11.2) the new number of training p ∈ [p for needing to update model parameter is determined using price differential proof methodmin,pmax];
11.3) new training sample is utilized, step 4) is repeated, to the relevant parameter Least-squares minimization of modelling by mechanism
Algorithm is updated;
11.4) new hybrid modeling in parallel will be obtained to apply in actual condition, obtains corresponding Accurate Prediction.
Technical concept of the invention are as follows: raw sample data collection is divided into training sample set and test sample collection first;
Mechanism model is established using training set, then multiple new training sample subsets, each subset of training are established with random method for resampling
Establish the base learner set of Elman neural network;Then clustering is carried out, the preferably individual conduct of performance in every cluster is chosen
Selective ensemble subset;The hybrid modeling method in parallel of the dynamic compensation mechanism is finally used for test sample collection, utilizes error
Judge whether to need to update mechanism model training sample and corresponding parameter according to similarity principle, obtains corresponding polypropylene melt
Melt exponential forecasting value.
Beneficial effects of the present invention are mainly manifested in:
1, random sampling with replacement is carried out by Bootstrap algorithm, establishes several new training samples with otherness
Subset, the contradiction between active balance " accuracy rate-otherness ", accuracy, generalization ability and the robust of lift scheme prediction
Property;
2, the submodel different using Elman neural network more meets and Nonlinear Dynamic in actual production process
Problem, lift scheme entirety predictive ability;
3, Fuzzy C-means (FCM) clustering algorithm provides a kind of new for the individual preferably link of selective ensemble model
Resolving ideas will screen individual submodel problem and be transformed to sample clustering problem, optimized individual submodel in every cluster is selected to represent
Entire classification can be made selective model being optimal of network, not only be reduced operation cost with exclusive segment redundancy individual,
Improve model generalization performance;
4, by calculating the similarity of current sample data and historical data, judge to update mechanism model relevant parameter must
The property wanted further enhances model entirety estimated performance, improves model stability;
5, the polypropylene melt index soft-sensing model established according to the present invention can preferably track the molten of production process
Melt index variation trend, there is preferable estimated performance and generalization ability, will optimize for polypropylene production operation and provide effectively
The quality control of polypropylene production process is effectively realized in technical support.
Detailed description of the invention
Fig. 1 is the double loop polypropylene production unit process schematic representations of certain petroleum chemical enterprise.
Fig. 2 is Elman neural network structure figure.
Fig. 3 is melt index model coefficient online updating implementation flow chart.
Fig. 4 is that the melt index on-line predictive model prediction result of mechanism model A compares figure.
Fig. 5 is the mixed model B melt index on-line predictive model prediction result of single Elman Neural Networks Error Compensation
Compare figure.
Fig. 6 is that the mixed model C melt index on-line predictive model of selective ensemble Elman Neural Networks Error Compensation is pre-
It surveys result and compares figure.
Fig. 7 is that the parallel hybrid model D melt index on-line predictive model prediction result of online updating strategy compares figure.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.
Referring to Fig.1~Fig. 7, a kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism, packet
Include following steps:
1) the double loop polypropylene production unit technique signals of certain petroleum chemical enterprise are as shown in Figure 1, in detailed analysis propylene polymerization
On the basis of dynamics and the bis- loop technology features of Spheripol, suitable auxiliary variable and leading variable are selected.Polypropylene melt
The auxiliary variable for melting index hard measurement includes the first endless tube density of hydrogen H21(ppm), the second endless tube density of hydrogen H22(ppm), it urges
Agent flow rate Ccat(kg/h), the first endless tube propylene monomer flow C31(t/h), the second endless tube propylene monomer flow C32(t/h),
One annular-pipe reactor temperature T1(DEG C), the second annular-pipe reactor temperature T2(DEG C), the first endless tube interlayer water temperature T11(DEG C), the second ring
Pipe clamp layer water temperature T22(℃).Leading variable is polypropylene melt index MI.Pass through Distributed Control System and experimental analysis database
Production operation data are acquired, raw sample data collection X is established;
2) abnormal data in the original sample collection X that step 1) obtains is rejected using Lai Yite criterion, is disappeared using filtering method
Except the random error in data, and data normalized is carried out, eliminates data dimension and difference in size brings model training
Influence;
3) according to setting ratio, the pretreated raw sample data collection X of step 2) is divided into training sample set XMWith
Test sample collection XK.100 groups are respectively extracted in two trade mark polypropylene data samples, wherein 140 groups of training sample, is left 60 groups
As sample to be predicted;
4) training sample X is utilizedMMiddle auxiliary variable and leading variable are established according to the mechanism knowledge of polypropylene production process
Mechanism model:
In formula, α0,···,α6It indicates model constants undetermined, recognizes model parameter with Least-squares minimization algorithm, obtain
It is as shown in table 1 to parameter estimation result:
α0 | α1 | α2 | α4 | α3 | α5 | α6 |
-28.6219 | -59.4126 | 1.60367 | -1.8465 | -0.564 | 1.296 | -0.16639 |
Table 15) by training sample set XMAuxiliary variable and step 4) in the obtained error of modelling by mechanism respectively as number
According to outputting and inputting for driving error compensation model, new training sample S is formed;
6) to training sample S, extraction yield p and resampling number c is determined, is repeated at random using Bootstrap algorithm
Sampling, establishes several training sample subsets { S1, S2, S3..., Sc, treatment process is as follows:
It 6.1) will be between modelling by mechanism obtains in the experimental analysis value after polypropylene off-line analysis and step 4 predicted value
Output vector of the error as data-driven error compensation model;
6.2) original training sample collection X is chosenMMiddle auxiliary variable yi=[yi,1,yi,2,…,yi,m]T∈RmIt is driven as data
The input of dynamic error compensation model;
6.3) to new training sample S, determine that extraction yield is 80%, resampling number 10, using Bootstrap algorithm into
Row random sampling with replacement establishes 10 training sample subset { S1, S2, S3..., S10};
7) 10 different Elman neural network submodels are established based on each sample set stand-alone training, treatment process is such as
Under:
7.1) to the training sample subset { S provided in step 6)1, S2, S3..., S10, Elman neural network structure is as schemed
Shown in 2, using 3-tier architecture, hidden layer node value range [7,11], and tansing type and s shape function are chosen respectively as hidden
Activation primitive containing layer;
7.2) for Elman neural network,Respectively input layer to hidden layer, accept layer to hidden layer connection weight
Weight matrix, f () indicate the excitation function of output unit, then corresponding hidden layer output are as follows:
The output for accepting layer indicates are as follows:
In formula, t indicates the moment;
7.3) the nonlinear state equation of output layer indicates are as follows:
In formula,Indicate connection weight matrix of the hidden layer to output layer, the excitation function of the implicit layer unit of g () expression;
8) Fuzzy C-means (FCM) clustering algorithm is used, determines that clusters number is 3, rationally screens in every cluster accuracy most
Alternatively property integrates submodel for integrating to good polypropylene melt index model subjects, and treatment process is as follows:
8.1) it is numbered using 10 polypropylene melt index Elman neural network submodels after training, and to instruction
The individual for practicing sample set is predicted to obtain prediction output matrix H;
8.2) using Fuzzy C-means (FCM) clustering algorithm to prediction output matrix H clustering.Its objective function indicates
Are as follows:
In formula, m is clusters number;N is sample number;B is fuzzy clustering index, generally takes b >=1;υj(j=1,2 ..., m)
The center clustered for j-th;uj(xi) indicate i-th of sample for jth class degree of membership, and meet:
8.3) the cluster centre matrix V=[υ obtainedj] and subordinated-degree matrix U=[uj(xi)] iterative formula be respectively as follows:
8.4) it by obtained subordinated-degree matrix U, obtains each Elman neural network and angle value is subordinate to every cluster;
8.5) each Elman neural network submodel is belonged in the maximum cluster of degree of membership, and according to average relative error
Minimum principle chooses the higher Elman neural network submodel of the accuracy alternatively integrated subset of property from every cluster.
The concrete outcome of polypropylene melt index submodel is screened such as using fuzzy clustering selection algorithm according to step 8)
Shown in table 2.
Cluster is other | Submodel number | Average relative error | Best submodel |
1 | 1,4,5,9 | 0.0894,0.1109,0.1087,0.1165 | 1 |
2 | 2,7,8 | 0.0499,0.0487,0.0529 | 7 |
3 | 3,6,10 | 0.0913,0.0614,0.0786 | 6 |
Table 29) setting permission estimated bias lower limit, it is good using learning training in new polypropylene production status
The evaluated error of Elman Neural Networks Error Compensation model prediction mechanism model and real system process, if error meets setting
Modelling by mechanism predicted value, then be added, the estimated value of direct output-parallel mixed model by value with error compensation model estimated value;
10) similarity for otherwise needing to calculate current production status sample data and historical data again, judges whether to meet
Sample Similarity lower limit 0.05 such as meets condition, then does not need to be added in training set, treatment process is as follows:
10.1) current production status sample data x is calculatednewWith history training sample data XMMiddle auxiliary variable data yi=
[yi,1,yi,2,…,yi,m]T∈RmThe formula of distance and angle is respectively as follows:
10.2) according to distance and angle, it can get the similarity of current production status sample data and historical data,
Calculation formula are as follows:
Snew,i=β exp (- l (xnew,yi))+(1-β)cosθ(xnew,yi)
In formula, β indicates weight coefficient and 0≤β≤1, β=0.95.
10.3) if the similarity obtained meets given Sample Similarity lower limit 0.1, do not need to say that it is added to training
It concentrates;
If 11) be unsatisfactory for the condition 0.1 of Sample Similarity lower limit, need to determine more new model using cross-validation method
The new number of training of parameter, and step 4) is repeated, mechanism model relevant parameter is updated, realizes known operation item
The Accurate Prediction of polypropylene melt index under part.Real-time update process is as shown in figure 3, treatment process is as follows:
11.1) if similarity obtained in step 10 is unsatisfactory for given Sample Similarity lower limit 0.1, need to update instruction
Practice sample number;
11.2) the new number of training p ∈ [p for needing to update model parameter is determined using price differential proof methodmin,pmax],
[pmin,pmax]=[15,30];
11.3) new training sample is utilized, step 4) is repeated, to the relevant parameter Least-squares minimization of modelling by mechanism
Algorithm is updated;
11.4) new hybrid modeling in parallel will be obtained to apply in actual condition, obtains corresponding Accurate Prediction.
Hybrid modeling method is compensated in order to preferably assess the dynamic that the present invention establishes, it is pre- online with polypropylene melt index
Verifying model prediction performance is offered, the mechanism model A based on single identification of Model Parameters is established respectively, is based on single Elman mind
Mixed model B, the mixed model C based on selective ensemble Elman Neural Networks Error Compensation and sheet through network error compensation
The parallel hybrid model D based on online updating strategy that text proposes, the four kinds of polypropylene melt indexes provided shown in following figure 4- Fig. 7
The comparison result of the predicted value off-line analysis value of model, the Zuo Bantu in each figure is trade mark I, and right half figure is trade mark II.
In order to which the estimated performance to model A, B, C and D carries out quantitative assessment, using average relative error (MRE), averagely absolutely
To error (MAE), root-mean-square error (RMSE) and related coefficient (R2) it is used as quantitatively evaluating index, calculation formula is as follows:
In formula, yi、The respectively off-line analysis value and model predication value of polypropylene melt index.
The qualitative assessment achievement data of 4 kinds of models is as shown in table 3.
Model | MRE | MAE(g*10min-1) | RMS(g*10min-1) | R2 |
A | 0.0947 | 0.2206 | 0.2817 | 0.9029 |
B | 0.0781 | 0.1583 | 0.2078 | 0.9448 |
C | 0.0622 | 0.1288 | 0.1892 | 0.9586 |
D | 0.0329 | 0.0788 | 0.1031 | 0.9879 |
Table 3
In conjunction with Fig. 4~Fig. 7 and table 3 it is found that the estimated performance of model A, B, C, D successively enhance.Model D is on two trades mark
Every estimated performance index it is optimal.
The model D established using the method for the present invention is had the advantage that through current historical sample through Elman nerve net
Network deviation compensation model obtains estimated bias, to assess whether device performance variable occurs significant change, if need to update
Model parameter selectively updates the unknowm coefficient in identification model, fully considers the mechanism model limited range established offline
With model On-line matching, to enhance model prediction accuracy and stability to a certain extent.It is raw that polypropylene can preferably be tracked
The melt index change trend of production process further improves the estimated performance and generalization ability of melt index soft-sensing model,
It ensure that the reliability and safety of polypropylene quality control in the actual production process.
Finally it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations, to the greatest extent
Present invention has been described in detail with reference to the aforementioned embodiments for pipe, those skilled in the art should understand that it still may be used
To modify to technical solution documented by previous embodiment, or some or all of the technical features are equal
Replacement and these are modified or replaceed, model defined by the claims in the present invention that it does not separate the essence of the corresponding technical solution
It encloses.
Claims (6)
1. a kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism, which is characterized in that the side
Method the following steps are included:
1) according to actual industrial production operating condition, auxiliary variable and leading variable is selected, Distributed Control System and experimental analysis are passed through
Database acquires production operation data, establishes raw sample data collection X;
2) abnormal data in the original sample collection X that step 1) obtains is rejected using Lai Yite criterion, number is eliminated using filtering method
Random error in, and data normalized is carried out, data dimension and difference in size are eliminated to model training bring shadow
It rings;
3) according to setting ratio, the pretreated raw sample data collection X of step 2) is divided into training sample set XMAnd test specimens
This collection XK;
4) training sample X is utilizedMMiddle auxiliary variable and leading variable establish mechanism model according to the mechanism knowledge of production process, and
Model parameter is recognized with Least-squares minimization algorithm;
5) by training sample set XMAuxiliary variable and the obtained error of modelling by mechanism respectively as data-driven error compensation mould
Type is output and input, and forms new training sample S;
6) to training sample S, extraction yield p and resampling number c are determined, random sampling with replacement is carried out using Bootstrap algorithm,
Establish several training sample subsets { S1, S2, S3..., Sc};
7) training parameter of Elman neural network is set, is trained based on various kinds book the set pair analysis model, establishes several
Elman neural network submodel;
8) Fuzzy C-Means Clustering Algorithm is used, determines clusters number, screens the best individual of accuracy in every cluster alternatively
Property integrated submodel for integrating;
9) setting allows estimated bias lower limit to utilize the good Elman nerve net of learning training in new polypropylene production status
Network error compensation model predicts the evaluated error of mechanism model and real system process, directly defeated if error meets setting value
The output valve of parallel hybrid model out;
10) similarity for otherwise needing to calculate current production status sample data and historical data again, judges whether to meet sample
Similarity lower limit such as meets condition, then does not need to be added in training set;
If 11) be unsatisfactory for the condition of Sample Similarity lower limit, need to determine the parameter of more new model using cross-validation method
New number of training, and step 4) is repeated, mechanism model relevant parameter is updated, is gathered under the conditions of realization known operation
The Accurate Prediction of propylene melt index.
2. a kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism according to claim 1,
It is characterized in that, the treatment process of the step 6) are as follows:
6.1) error between the predicted value for obtaining experimental analysis value and modelling by mechanism is as data-driven error compensation model
Output vector xi=[xi,1,xi,2,…,xi,n]T∈Rn;
6.2) original training sample collection X is chosenMMiddle auxiliary variable yi=[yi,1,yi,2,…,yi,m]T∈RmAs data-driven error
The input of compensation model;
6.3) to new training sampleDetermine that extraction yield is 80%, resampling number 10 uses
Bootstrap algorithm carries out random sampling with replacement, establishes 10 training sample subset { S1, S2, S3..., S10}。
3. a kind of polypropylene melt index hybrid modeling side based on dynamic error compensation mechanism according to claim 1 or claim 2
Method, which is characterized in that the treatment process of the step 7) are as follows:
7.1) to the training sample subset { S provided1, S2, S3..., Sc, Elman neural network uses 3-tier architecture, chooses implicit
Node layer number, and the activation primitive of tansing type and s shape function as hidden layer is chosen respectively;
7.2) for Elman neural network,Respectively input layer to hidden layer, accept layer to hidden layer connection weight square
Battle array, f () indicate the excitation function of output unit, then corresponding hidden layer output are as follows:
In formula,Respectively input layer indicates defeated to hidden layer, the connection weight matrix of undertaking layer to hidden layer, f ()
The excitation function of unit out.
The output for accepting layer indicates are as follows:
In formula, t indicates the moment,;Respectively input layer to hidden layer, accept layer to hidden layer connection weight square
Battle array, f () indicate the excitation function of output unit.
7.3) the nonlinear state equation of output layer indicates are as follows:
In formula,Indicate connection weight matrix of the hidden layer to output layer, the excitation function of the implicit layer unit of g () expression.
4. a kind of polypropylene melt index hybrid modeling side based on dynamic error compensation mechanism according to claim 1 or claim 2
Method, which is characterized in that the treatment process of the step 8) are as follows:
8.1) individual of training sample subset is predicted to obtain using the Elman neural network submodel after training predict it is defeated
Matrix H out;
8.2) using Fuzzy C-Means Clustering Algorithm to prediction output matrix H clustering, objective function is indicated are as follows:
In formula, mn is clusters number;N is sample number;B is fuzzy clustering index, takes b >=1;υj(j=1,2 ..., m) it is j-th
The center of cluster;uj(xi) indicate i-th of sample for jth class degree of membership, and meet:
8.3) the cluster centre matrix V=[υ obtainedj] and subordinated-degree matrix U=[uj(xi)] iterative formula be respectively as follows:
8.4) it by obtained subordinated-degree matrix U, obtains each Elman neural network and angle value is subordinate to every cluster;
8.5) each Elman neural network submodel is belonged in the maximum cluster of degree of membership, and according to average relative error minimum
Principle chooses the higher Elman neural network submodel of the accuracy alternatively integrated subset of property from every cluster.
5. a kind of polypropylene melt index hybrid modeling side based on dynamic error compensation mechanism according to claim 1 or claim 2
Method, which is characterized in that the treatment process of the step 10) are as follows:
10.1) current production status sample data x is calculatednewWith history training sample data XMMiddle auxiliary variable data yi=
[yi,1,yi,2,…,yi,m]T∈RmThe formula of distance and angle is respectively as follows:
In formula, l (xnew,xm) and θ (xnew,xm) respectively indicate the calculation formula of distance and angle;
10.2) according to distance and angle, the similarity of current production status sample data and historical data is obtained, is calculated public
Formula are as follows:
Snew,i=β exp (- l (xnew,yi))+(1-β)cosθ(xnew,yi)
In formula, β indicates weight coefficient and 0≤β≤1;
10.3) if the similarity obtained meets given Sample Similarity lower limit, do not need to say that it is added to training set sample
In.
6. a kind of polypropylene melt index hybrid modeling side based on dynamic error compensation mechanism according to claim 1 or claim 2
Method, which is characterized in that the treatment process of the step 11) are as follows:
11.1) if similarity obtained in step 10) is unsatisfactory for given Sample Similarity lower limit, need to update training sample
Number;
11.2) the new number of training p ∈ [p for needing to update model parameter is determined using price differential proof methodmin,pmax];
11.3) new training sample is utilized, step 4) is repeated, to the relevant parameter Least-squares minimization algorithm of modelling by mechanism
It is updated;
11.4) new hybrid modeling in parallel will be obtained to apply in actual condition, obtains corresponding Accurate Prediction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534700.1A CN110322933A (en) | 2019-06-20 | 2019-06-20 | A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910534700.1A CN110322933A (en) | 2019-06-20 | 2019-06-20 | A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110322933A true CN110322933A (en) | 2019-10-11 |
Family
ID=68120005
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910534700.1A Pending CN110322933A (en) | 2019-06-20 | 2019-06-20 | A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110322933A (en) |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111128312A (en) * | 2019-12-19 | 2020-05-08 | 湖南工业大学 | Zinc oxide volatilization kiln mixed modeling method based on mechanism and support vector machine |
CN111912875A (en) * | 2020-06-23 | 2020-11-10 | 宁波大学 | Fractionating tower benzene content soft measurement method based on stack type Elman neural network |
CN111914476A (en) * | 2020-06-23 | 2020-11-10 | 宁波大学 | Online soft measurement method for butane content of product at bottom of debutanizer |
CN112257341A (en) * | 2020-10-20 | 2021-01-22 | 浙江大学 | Customized product performance prediction method based on heterogeneous data difference compensation fusion |
CN112577161A (en) * | 2019-09-30 | 2021-03-30 | 北京国双科技有限公司 | Air conditioner energy consumption model training method and air conditioner system control method |
CN113065269A (en) * | 2021-04-02 | 2021-07-02 | 上海交通大学 | Composite material multi-scale concurrent simulation method and system |
CN113139155A (en) * | 2021-04-23 | 2021-07-20 | 南京富岛信息工程有限公司 | Large-range crude oil blending selection optimization method |
CN114596919A (en) * | 2022-05-10 | 2022-06-07 | 安徽瑞邦数科科技服务有限公司 | Index prediction method and system and application thereof in phosphoric acid production |
WO2022155787A1 (en) * | 2021-01-19 | 2022-07-28 | 深圳大学 | Extreme learning machine training method, extreme learning machine training apparatus, and terminal device |
CN117055487A (en) * | 2023-08-24 | 2023-11-14 | 北京科技大学 | Two-roller skew rolling perforation parameter optimization method based on mechanism data hybrid driving |
CN117150425A (en) * | 2023-07-10 | 2023-12-01 | 郑州轻工业大学 | Segment erector motion state prediction method based on mechanism data fusion |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5675253A (en) * | 1991-11-20 | 1997-10-07 | Auburn International, Inc. | Partial least square regression techniques in obtaining measurements of one or more polymer properties with an on-line nmr system |
CN101315557A (en) * | 2008-06-25 | 2008-12-03 | 浙江大学 | Propylene polymerization production process optimal soft survey instrument and method based on genetic algorithm optimization BP neural network |
CN103267826A (en) * | 2013-04-17 | 2013-08-28 | 沈阳大学 | Soft measurement method for online detection of gelatin concentration |
CN103294601A (en) * | 2013-07-03 | 2013-09-11 | 中国石油大学(华东) | Software reliability forecasting method based on selective dynamic weight neural network integration |
CN104899656A (en) * | 2015-06-05 | 2015-09-09 | 三峡大学 | Wind power combined predication method based on ensemble average empirical mode decomposition and improved Elman neural network |
CN105243461A (en) * | 2015-11-20 | 2016-01-13 | 江苏省电力公司 | Short-term load forecasting method based on artificial neural network improved training strategy |
CN106168759A (en) * | 2016-07-12 | 2016-11-30 | 武汉长江仪器自动化研究所有限公司 | A kind of coagulant dosage control method and system based on artificial neural network algorithm |
CN108053052A (en) * | 2017-11-20 | 2018-05-18 | 淮阴工学院 | A kind of oil truck oil and gas leakage speed intelligent monitor system |
CN108093085A (en) * | 2018-01-16 | 2018-05-29 | 成都寻道科技有限公司 | A kind of big data system real time service parameter adaptive adjusting method |
CN108985574A (en) * | 2018-06-23 | 2018-12-11 | 浙江工业大学 | A kind of polypropylene melt index flexible measurement method based on selective ensemble extreme learning machine |
-
2019
- 2019-06-20 CN CN201910534700.1A patent/CN110322933A/en active Pending
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5675253A (en) * | 1991-11-20 | 1997-10-07 | Auburn International, Inc. | Partial least square regression techniques in obtaining measurements of one or more polymer properties with an on-line nmr system |
CN101315557A (en) * | 2008-06-25 | 2008-12-03 | 浙江大学 | Propylene polymerization production process optimal soft survey instrument and method based on genetic algorithm optimization BP neural network |
CN103267826A (en) * | 2013-04-17 | 2013-08-28 | 沈阳大学 | Soft measurement method for online detection of gelatin concentration |
CN103294601A (en) * | 2013-07-03 | 2013-09-11 | 中国石油大学(华东) | Software reliability forecasting method based on selective dynamic weight neural network integration |
CN104899656A (en) * | 2015-06-05 | 2015-09-09 | 三峡大学 | Wind power combined predication method based on ensemble average empirical mode decomposition and improved Elman neural network |
CN105243461A (en) * | 2015-11-20 | 2016-01-13 | 江苏省电力公司 | Short-term load forecasting method based on artificial neural network improved training strategy |
CN106168759A (en) * | 2016-07-12 | 2016-11-30 | 武汉长江仪器自动化研究所有限公司 | A kind of coagulant dosage control method and system based on artificial neural network algorithm |
CN108053052A (en) * | 2017-11-20 | 2018-05-18 | 淮阴工学院 | A kind of oil truck oil and gas leakage speed intelligent monitor system |
CN108093085A (en) * | 2018-01-16 | 2018-05-29 | 成都寻道科技有限公司 | A kind of big data system real time service parameter adaptive adjusting method |
CN108985574A (en) * | 2018-06-23 | 2018-12-11 | 浙江工业大学 | A kind of polypropylene melt index flexible measurement method based on selective ensemble extreme learning machine |
Non-Patent Citations (5)
Title |
---|
HAICHUAN LOU等: "Inferential Model for Industrial Polypropylene Melt Index Prediction with Embedded Priori Knowledge and Delay Estimation", 《I&EC RESEARCH》 * |
LUYUE XIA 等: "Inferential Estimation of Polypropylene Melt Index Using Stacked Neural Networks Based on Absolute Error Criteria", 《2010 INTERNATIONAL CONFERENCE ON COMPUTER》 * |
MIN ZHANG 等: "Multi-model quality prediction approach using fuzzy C-means clustering and support vector regression", 《ADVANCES IN MECHANICAL ENGINEERING》 * |
YING ZHANG等: "An improved Elman neural network with piecewise weighted gradient__for time series prediction", 《NEUROCOMPUTING》 * |
陈雯柏 等: "《人工神经网络原理与实践》", 31 January 2016 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112577161A (en) * | 2019-09-30 | 2021-03-30 | 北京国双科技有限公司 | Air conditioner energy consumption model training method and air conditioner system control method |
CN111128312A (en) * | 2019-12-19 | 2020-05-08 | 湖南工业大学 | Zinc oxide volatilization kiln mixed modeling method based on mechanism and support vector machine |
CN111128312B (en) * | 2019-12-19 | 2022-06-17 | 湖南工业大学 | Zinc oxide volatilization kiln mixed modeling method based on mechanism and support vector machine |
CN111914476B (en) * | 2020-06-23 | 2023-09-29 | 宁波大学 | On-line soft measurement method for butane content of debutanizer bottom product |
CN111912875B (en) * | 2020-06-23 | 2024-02-13 | 江苏淮河化工有限公司 | Fractionation tower benzene content soft measurement method based on stacked Elman neural network |
CN111912875A (en) * | 2020-06-23 | 2020-11-10 | 宁波大学 | Fractionating tower benzene content soft measurement method based on stack type Elman neural network |
CN111914476A (en) * | 2020-06-23 | 2020-11-10 | 宁波大学 | Online soft measurement method for butane content of product at bottom of debutanizer |
CN112257341A (en) * | 2020-10-20 | 2021-01-22 | 浙江大学 | Customized product performance prediction method based on heterogeneous data difference compensation fusion |
CN112257341B (en) * | 2020-10-20 | 2022-04-26 | 浙江大学 | Customized product performance prediction method based on heterogeneous data difference compensation fusion |
WO2022155787A1 (en) * | 2021-01-19 | 2022-07-28 | 深圳大学 | Extreme learning machine training method, extreme learning machine training apparatus, and terminal device |
CN113065269B (en) * | 2021-04-02 | 2022-06-28 | 上海交通大学 | Composite material multi-scale concurrent simulation method and system |
CN113065269A (en) * | 2021-04-02 | 2021-07-02 | 上海交通大学 | Composite material multi-scale concurrent simulation method and system |
CN113139155A (en) * | 2021-04-23 | 2021-07-20 | 南京富岛信息工程有限公司 | Large-range crude oil blending selection optimization method |
CN113139155B (en) * | 2021-04-23 | 2024-02-06 | 南京富岛信息工程有限公司 | Large-scale crude oil blending selection optimization method |
CN114596919A (en) * | 2022-05-10 | 2022-06-07 | 安徽瑞邦数科科技服务有限公司 | Index prediction method and system and application thereof in phosphoric acid production |
CN117150425A (en) * | 2023-07-10 | 2023-12-01 | 郑州轻工业大学 | Segment erector motion state prediction method based on mechanism data fusion |
CN117150425B (en) * | 2023-07-10 | 2024-04-26 | 郑州轻工业大学 | Segment erector motion state prediction method based on mechanism data fusion |
CN117055487A (en) * | 2023-08-24 | 2023-11-14 | 北京科技大学 | Two-roller skew rolling perforation parameter optimization method based on mechanism data hybrid driving |
CN117055487B (en) * | 2023-08-24 | 2024-04-16 | 北京科技大学 | Two-roller skew rolling perforation parameter optimization method based on mechanism data hybrid driving |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110322933A (en) | A kind of polypropylene melt index hybrid modeling method based on dynamic error compensation mechanism | |
Newhart et al. | Data-driven performance analyses of wastewater treatment plants: A review | |
Xiong et al. | JITL based MWGPR soft sensor for multi-mode process with dual-updating strategy | |
Chen et al. | Modelling and optimization of fed-batch fermentation processes using dynamic neural networks and genetic algorithms | |
Pan et al. | Just-in-time learning based soft sensor with variable selection and weighting optimized by evolutionary optimization for quality prediction of nonlinear processes | |
Zhang et al. | Double-level locally weighted extreme learning machine for soft sensor modeling of complex nonlinear industrial processes | |
CN105654102A (en) | Data processing device and data processing method | |
CN108985574A (en) | A kind of polypropylene melt index flexible measurement method based on selective ensemble extreme learning machine | |
CN105868164B (en) | A kind of soft-measuring modeling method based on the linear dynamic system model for having supervision | |
CN103226728B (en) | High density polyethylene polymerization cascade course of reaction Intelligent Measurement and yield optimization method | |
Li et al. | Gaussian process regression with heteroscedastic noises—A machine-learning predictive variance approach | |
CN111027733A (en) | Petrochemical device product yield optimization method based on big data technology | |
Hu et al. | Weighted kernel fuzzy C-means-based broad learning model for time-series prediction of carbon efficiency in iron ore sintering process | |
CN100580585C (en) | Soft measuring method of industrial process under condition of small sample | |
Yeganeh et al. | Using evolutionary artificial neural networks in monitoring binary and polytomous logistic profiles | |
CN110378035A (en) | It is a kind of that soft-measuring modeling method is hydrocracked based on deep learning | |
Yasmin et al. | Improved support vector machine using optimization techniques for an aerobic granular sludge | |
sheng Wang et al. | D-FNN based soft-sensor modeling and migration reconfiguration of polymerizing process | |
CN109493921A (en) | A kind of atmospheric distillation process modeling approach based on multi-agent system model | |
CN111204867B (en) | Membrane bioreactor-MBR membrane pollution intelligent decision-making method | |
CN201330211Y (en) | Working parameter self-optimizing simulation system for sewage treatment plant | |
Elfelly et al. | A multimodel approach for complex systems modeling based on classification algorithms | |
Liu et al. | Real-time optimization compensation method based on a novel two-level multi-block hybrid model for the hydrometallurgy process | |
Han et al. | Piecewise sliding-mode control for sludge bulking under multiple operating conditions | |
Mahboob et al. | Quality enhancement at higher education institutions by early identifying students at risk using data mining |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191011 |