CN109543916A - Silicon rod growth rate prediction model in a kind of polycrystalline silicon reducing furnace - Google Patents
Silicon rod growth rate prediction model in a kind of polycrystalline silicon reducing furnace Download PDFInfo
- Publication number
- CN109543916A CN109543916A CN201811454413.1A CN201811454413A CN109543916A CN 109543916 A CN109543916 A CN 109543916A CN 201811454413 A CN201811454413 A CN 201811454413A CN 109543916 A CN109543916 A CN 109543916A
- Authority
- CN
- China
- Prior art keywords
- data
- value
- sample
- growth rate
- predicted
- 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.)
- Granted
Links
- XUIMIQQOPSSXEZ-UHFFFAOYSA-N Silicon Chemical compound [Si] XUIMIQQOPSSXEZ-UHFFFAOYSA-N 0.000 title claims abstract description 48
- 229910052710 silicon Inorganic materials 0.000 title claims abstract description 48
- 239000010703 silicon Substances 0.000 title claims abstract description 48
- 229910021420 polycrystalline silicon Inorganic materials 0.000 title claims abstract description 24
- 239000013598 vector Substances 0.000 claims abstract description 28
- 238000004519 manufacturing process Methods 0.000 claims abstract description 23
- 238000012549 training Methods 0.000 claims abstract description 18
- 238000012360 testing method Methods 0.000 claims abstract description 14
- 238000012797 qualification Methods 0.000 claims abstract description 11
- 238000012216 screening Methods 0.000 claims abstract description 8
- 238000011156 evaluation Methods 0.000 claims abstract description 6
- 238000003066 decision tree Methods 0.000 claims abstract description 5
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 230000001737 promoting effect Effects 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 11
- 229910052739 hydrogen Inorganic materials 0.000 claims description 6
- 239000001257 hydrogen Substances 0.000 claims description 6
- UFHFLCQGNIYNRP-UHFFFAOYSA-N Hydrogen Chemical compound [H][H] UFHFLCQGNIYNRP-UHFFFAOYSA-N 0.000 claims description 5
- ZDHXKXAHOVTTAH-UHFFFAOYSA-N trichlorosilane Chemical compound Cl[SiH](Cl)Cl ZDHXKXAHOVTTAH-UHFFFAOYSA-N 0.000 claims description 4
- 239000005052 trichlorosilane Substances 0.000 claims description 4
- 230000005484 gravity Effects 0.000 claims description 3
- 238000007619 statistical method Methods 0.000 claims description 3
- 241001269238 Data Species 0.000 claims description 2
- 230000008901 benefit Effects 0.000 claims description 2
- 238000006243 chemical reaction Methods 0.000 abstract description 11
- 238000011160 research Methods 0.000 abstract description 3
- 229920005591 polysilicon Polymers 0.000 description 8
- 238000000034 method Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 4
- 230000009467 reduction Effects 0.000 description 4
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000005265 energy consumption Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 2
- 241001347978 Major minor Species 0.000 description 1
- 240000007594 Oryza sativa Species 0.000 description 1
- 235000007164 Oryza sativa Nutrition 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 150000002431 hydrogen Chemical class 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 239000003208 petroleum Substances 0.000 description 1
- 239000002994 raw material Substances 0.000 description 1
- 235000009566 rice Nutrition 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/04—Manufacturing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Business, Economics & Management (AREA)
- General Health & Medical Sciences (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Operations Research (AREA)
- Entrepreneurship & Innovation (AREA)
- Evolutionary Computation (AREA)
- Quality & Reliability (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Manufacturing & Machinery (AREA)
- Primary Health Care (AREA)
- Silicon Compounds (AREA)
Abstract
The invention discloses silicon rod growth rate prediction models in a kind of polycrystalline silicon reducing furnace, including data acquisition and preprocessing module, for acquiring data relevant to silicon rod growth rate by sensor, and carry out data cleansing to the data of acquisition；Data set screening module obtains training set and test set for screening to data set of the data acquisition module after data acquisition, data cleansing；Trained and evaluation module is trained for promoting decision-tree model to gradient, is assessed using average relative error and qualification rate model, module to be predicted, for inputting feature vector to be predicted；Prediction result module exports prediction result for the feature vector to be predicted to be input in prediction model.The present invention has abandoned influence of the complex chemical reaction to growth rate in traditional theoretical research, without considering complicated chemical reaction, realizes and estimates to polycrystalline silicon growth rate, to improve production efficiency, shorten production cycle offer guidance.
Description
Technical field
The present invention relates to polysilicon production process technical fields, and in particular to one kind improves GBDT based on grey association
Silicon rod growth rate prediction model in the polycrystalline silicon reducing furnace of regression algorithm.
Background technique
Polysilicon is widely used in photovoltaic, in the equipment such as electronic device.It is increasingly depleted with energy such as petroleum, too
The clean energy resourcies such as positive energy are widely paid close attention to, and the raw materials requirement based on polysilicon is caused to be good at increasingly.Production polysilicon at present
Technique in, it is most popular to belong to improved Siemens technique, account for 80% or so of whole world silicon total amount.The main life in China
The producer for producing polysilicon also mostly produces polysilicon using improved Siemens.But the technique is also along with energy loss simultaneously
Greatly, the disadvantages of production cycle is long.
And for current 24 domestic stick reduction furnaces, the time that silicon rod grows into required diameter completely probably needs 120
Hour or so, the production cycle is too long to limit the yield of production, while also along with the increase of power consumption.The production time of polysilicon
It is mainly related with the growth rate of silicon rod, but the growth rate of silicon rod is subject to each technological parameter, and coupling between parameter
Group photo is rung；The major-minor chemical reaction being related to inside reduction furnace simultaneously has upper ten, influences each other between each chemical reaction, reaction is dynamic
Mechanics is complicated.Reaction rate is speculated from reaction mechanism, and mainly reaction power is studied from theory on Math at present,
For live real-time, operability is not strong.The quality that too fast growth rate may cause silicon rod cannot be guaranteed, and be easy
Occur " puffed rice ", excessively slow growth rate again limits the production capacity of reduction furnace, increases energy consumption；Therefore to polycrystalline silicon growth rate
It is studied, improves growth rate under the premise of guaranteeing quality, growth rate is assessed, improve production efficiency, reduced
Energy consumption is of great importance.
Most enterprises mainly pass through visor to the growth of silicon rod at present or detection device carries out silicon rod in observation furnace
Growing state, make corresponding adjusting parameter again after obtaining the partial picture in furnace, there are lag issues；And to most of silicon
The research of stick Growth rate hypothesis is not strong for live operability, and reaction is complicated, be merely able to it is online under analyzed, can not needle
Scene is assessed.
Summary of the invention
It is influenced for silicon rod growth rate in polycrystalline silicon reducing furnace by Multi-parameter coupling problem, the present invention establishes polysilicon
The prediction model of silicon rod growth rate in reduction furnace obtains the relationship and changing rule between technological parameter and growth rate；Benefit
Technological parameter is selected with the prediction model, to improve growth rate, shortens the production cycle, production energy consumption offer is provided and is referred to
Lead foundation.
In order to realize above-mentioned task, the invention adopts the following technical scheme:
Silicon rod growth rate prediction model in a kind of polycrystalline silicon reducing furnace, comprises the following modules:
Data acquisition and preprocessing module, for acquiring data relevant to silicon rod growth rate, building by sensor
Data set, and data cleansing is carried out to the data of acquisition；
Data set screening module, for being carried out to data set of the data acquisition module after data acquisition, data cleansing
Screening, obtains training set and test set；
Trained and evaluation module is trained for promoting decision-tree model to gradient, using average relative error and
Qualification rate assesses model, obtains prediction model；It specifically includes:
(1) training set is inputted, sample number is M in training set；The number of iterations T and loss function L is set；The loss letter
Number uses square error loss function,Wherein y indicates that the true value of sample, y ' (x) indicate
Predicted value；
(2) weak learner is initialized
Find the constant value c for minimizing loss function:
In above formula, f0It (X) is weak learner, L () indicates loss function, yiIndicate the true output of i-th of sample；
(3) start iteration, in total iteration T times
(3-1) calculates negative gradient
The negative gradient of loss function when calculating each iteration: rti=yi-ft-1(Xi)；T=1,2...T, i=1,2...M,
ft-1(Xi) indicate the obtained strong learner of the t-1 times iteration to sample XiPredicted value；
(3-2) utilizes (Xi,rti) the weak regression tree of fitting, obtain the t weak regression tree；Its corresponding leaf node region is
Rtj, wherein j=1,2...J, J are the leaf node number of the t weak regression tree, t=1,2 ... T；
(3-3) is directed to leaf node region Rtj, find the best output valve c of fitting leaf nodetj:
(3-4) updates strong learner:
Wherein, ft-1(X) be the t-1 times iteration when strong learner, ftIt (X) is updated strong learner, the value of v
Range is 0 < v≤1, I expression indicator function, works as Xi∈RtjWhen, otherwise value 1 is 0；
(4) iteration terminates to obtain final strong learner:
Wherein, fT(X) the strong learner to be obtained after T iteration；
(5) strong learner is assessed using test set, obtains prediction model；
Module to be predicted, for inputting feature vector to be predicted；
Prediction result module exports prediction result for the feature vector to be predicted to be input in prediction model.
Further, the related data of the acquisition, comprising: furnace pressure, intake velocity, the amounts of hydrogen and trichlorine of charging
The amount of hydrogen silicon, silicon rod surface temperature, intake air temperature, the diameter of silicon rod, historical growth rate, silicon rod logarithm, furnace wall temperature；It will
One group of these data is as a sample.
Further, the data to acquisition carry out data cleansing, comprising:
Remember that the sample size acquired in total is m, for each of data set sample, as there are missing numbers in sample
According to, then by the missing data using the previous production cycle, the average value of corresponding data replaces in the latter production cycle, with into
The processing of row missing values；For outlier processing, using statistical method, the exception of every a kind of data is filtered out from data set
Value, is then handled using the exceptional value as missing values, or corrected with the average value of every other same type data.
It is further, described that data set of the data acquisition module after data acquisition, data cleansing is screened,
Obtain training set and test set, comprising:
The data in each sample concentrated to data are standardized, formula are as follows:
In above formula, xijIndicate the value of the jth class influence factor in i-th of sample, x 'ijIndicate xijBy standardization
Value afterwards, max (xj)、min(xj) respectively indicate maximum value, the minimum value of all jth class influence factors in data set；Wherein, i
=1,2 ... m；J=1,2 ... n, m indicate that sample size, n are the number of influence factor in sample；
Calculate the specific gravity p of j-th of influence factor in i-th of sampleij:
Calculate the entropy e of j-th of influence factorj:
In above formula,
Calculate the entropy weight w of j-th of influence factorj:
To obtain the weight vectors of each influence factor:
W=[w1 w2 w3 ... wn]
Calculate the initial matrix of each influence factor:
To initial matrix standardization processing:
Wherein:Wherein i=1,2 ... m, j=1,2 ... n；
If the feature vector of period to be predicted is y0=[y01 y02 … y0j], then characteristics of time interval vector y to be predicted0With i-th
The feature vector y of a sampleiIt is a in the degree of association of j-th of influence factorij
Wherein, y0jIndicate j-th of influence factor value of characteristics of time interval vector to be predicted, λ indicates resolution ratio；
Then degree of association Matrix C are as follows:
Wherein: amnIndicate the association angle value of n-th of influence factor of m-th of sample；
Weighted association judgment matrix are as follows:
Then projection value are as follows:
Wherein EiFor projection value of i-th of sample in the feature vector of period to be predicted；
According to projection value according to being arranged from big to small, threshold value is set, selects all samples greater than threshold value as phase
Like the data set of period, and training set and test set are divided on the basis of this data set；
Further, described that strong learner is assessed using test set, obtain prediction model, comprising:
Calculate average relative error:
Wherein, K indicates the sum of the future position in predicted time section, yiIndicate i moment silicon rod growth in predicted time section
Rate true value, y 'iIndicate i moment growth rate predicted value in predicted time section；
Qualification rate:
Wherein, Q (e < 4%) indicates the number of point of the relative error less than 4%, and K indicates future position sum；
If final strong learner meets the average relative error and qualification rate of setting, then will finally strong learner be used as in advance
Survey model.
The present invention has following technical characterstic compared with prior art:
1. the present invention considers influence of a variety of coupling factors to growth rate, real without considering complicated chemical reaction
Show and polycrystalline silicon growth rate has been estimated；The present invention can carry out the adjusting and optimizing of technological parameter in advance, guarantee the quality of silicon rod
Shorten production cycle offer guidance with growth rate to improve production efficiency.
2. scene specific aim strong real-time of the invention, relative to polynomial regression model, can handle various types with it is non-thread
Property data, strong robustness, accuracy rate is high.
3. the present invention has abandoned influence of the complex chemical reaction to growth rate in traditional theoretical research, the side of estimating is utilized
Case more succinctly easily can obtain each influence factor to the relationship of growth rate, strong operability.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of prediction model of the invention；
Fig. 2 is the workflow schematic diagram of data set screening module；
Fig. 3 is trained and evaluation module workflow schematic diagram；
Fig. 4 is the schematic diagram of model training process.
Specific embodiment
Thinking of the invention is to establish the prediction model of the growth rate of silicon rod in polycrystalline silicon reducing furnace, constructs multiple inputs
(such as furnace pressure, the speed of air inlet, the inlet amount of hydrogen and trichlorosilane, the diameter etc. of silicon rod), it is single to export the (life of silicon rod
Long rate) prediction model, provide guidance for the coupled problem of Optimizing Process Parameters, obtain Gao Erwen while ensuring quality
Fixed growth rate improves production efficiency.
The invention discloses silicon rod growth rate prediction models in a kind of polycrystalline silicon reducing furnace, comprise the following modules:
1. data acquisition and preprocessing module
For acquiring data relevant to silicon rod growth rate by sensor, data set is constructed；And to the data of acquisition
Carry out data cleansing.Specifically:
The acquisition of 1.1 data
Acquiring the data relevant to silicon rod growth rate by sensor in this programme includes: furnace pressure (MPa),
Intake velocity (m/s), the amounts of hydrogen (m of charging3) with the amount (kg) of trichlorosilane, silicon rod surface temperature (K), intake air temperature (K),
The diameter (mm) of silicon rod, historical growth rate (mm/h), silicon rod logarithm, furnace wall temperature (K)；Using one group of these data as one
A sample, each element in sample is an influence factor.
1.2 data prediction
Since there may be missing values or exceptional values for the data that acquire from sensor, missing values and exception need to be carried out
Value processing.
Missing values and outlier processing:
Remember that the sample size that acquires in total is m, for each of data set sample (i.e. the data at some moment),
If there are missing datas in sample, then the missing data is corresponded into number in the production cycle using previous production cycle, the latter
According to average value replace, to carry out missing values processing；Such as the intake velocity missing in a certain group of data, then with previous production
The average value of intake velocity replaces in period, the latter production cycle.For outlier processing, using statistical method, example
It, then can be using the exceptional value as missing if 3 σ principles of normal distribution filter out the exceptional value of every a kind of data from data set
Value is handled, or is corrected with the average value of every other same type data.
2. data set screening module
For screening to data set of the data acquisition module after data acquisition, data cleansing, training set is obtained
And test set, specifically:
The data in each sample that 2.1 pairs of data are concentrated are standardized, formula are as follows:
In above formula, xijIndicate the value of j-th of influence factor in i-th of sample, x 'ijIndicate xijBy standardization
Value afterwards, max (xj)、min(xj) respectively indicate maximum value, the minimum value of all jth class influence factors in data set.Wherein, i
=1,2 ... m；J=1,2 ... n, m indicate that sample size, n are the number of influence factor in sample.
Calculate the specific gravity p of j-th of influence factor in i-th of sampleij:
In above formula, m indicates sample size；
Calculate the entropy e of j-th of influence factorj:
In above formula,
Calculate the entropy weight w of j-th of influence factorj:
Wherein, j=1,2 ... n.
To obtain the weight vectors of each influence factor:
W=[w1 w2 w3 ... wn]
2.2 incidence matrix
Calculate the initial matrix of each influence factor:
To initial matrix standardization processing:
Wherein:Wherein i=1,2 ... m, j=1,2 ... n.
Assuming that the feature vector of period to be predicted is y0=[y01 y02 ... y0j], then characteristics of time interval vector y to be predicted0With
The feature vector y of i-th of sampleiIt is a in the degree of association of j-th of influence factorij:
Wherein, y0jIndicate j-th of influence factor value of characteristics of time interval vector to be predicted, λ indicates resolution ratio, generally takes
0.5。
Then degree of association Matrix C are as follows:
Wherein: amnIndicate that the association angle value of n-th of influence factor of m-th of sample, the first row indicate optimal situation,
Therefore the first row element is all 1.
Weighted association judgment matrix are as follows:
Then projection value are as follows:
Wherein EiFor projection value of i-th of sample in the feature vector of period to be predicted, i=1,2 ... m
2.3 according to projection value according to being arranged from big to small, threshold value is set, select all samples greater than threshold value as
The data set of similar period, and training set and test set are divided on the basis of this data set.
3. trained and evaluation module
It is instructed for promoting decision tree GBDT (Gradient Boosting Decision Tree) model to gradient
Practice, model assessed using average relative error and qualification rate, the specific steps are as follows:
3.1 input training sets, and sample number is M in training set；The number of iterations T and loss function L is set；The loss letter
Number uses square error loss function,Wherein y indicates that the true value of sample, y ' (x) indicate
Predicted value, L () are loss function.
The 3.2 weak learners of initialization
Find the constant value c for minimizing loss function:
In above formula, f0It (X) is weak learner, L () indicates loss function, yiIndicate the true output of i-th of sample,
That is the true growth rate of silicon rod；C initialization value is generally the average value of all training set data y.
3.3 start iteration, in total iteration T times
3.3.1 negative gradient is calculated
The negative gradient of loss function is obtained due to using square error as loss function when calculating each iteration
Negative gradient is also referred to as regression criterion rti=yi-ft-1(Xi)；T=1,2...T, i=1,2...M, ft-1(Xi) indicate to change for the t-1 times
The strong learner that generation obtains is to sample XiPredicted value；
3.3.2 utilizing (Xi,rti) the weak regression tree of fitting, obtain the t weak regression tree；Its corresponding leaf node region is
Rtj, wherein j=1,2...J, J are the leaf node number of the t weak regression tree, t=1,2 ... T；
3.3.3 it is directed to leaf node region Rtj, find the best output valve c of fitting leaf nodetj:
3.3.4 strong learner is updated:
Wherein, ft-1(X) be the t-1 times iteration when strong learner, ftIt (X) is updated strong learner, the value of v
Range is 0 < v≤1, I expression indicator function, works as Xi∈RtjWhen, otherwise value 1 is 0；
3.4 iteration terminate to obtain final strong learner:
Wherein, fT(X) the strong learner to be obtained after T iteration；
3.5 assess strong learner using test set, and evaluation index is used to the good of judgment models using average relative error
It is bad, with the stability of yield analysis model:
Average relative error:
Wherein, K indicates the sum of the future position in predicted time section, yiIndicate i moment silicon rod growth in predicted time section
Rate true value, y 'iIndicate i moment growth rate predicted value in predicted time section.
Qualification rate:
Wherein, Q (e < 4%) indicates the number of point of the relative error less than 4%, and K indicates future position sum.
If final strong learner meets the average relative error and qualification rate of setting, then will finally strong learner be used as in advance
Survey model；It is such as unsatisfactory for error and qualification rate, then re-start training and then is judged again.
4. module to be predicted
For inputting feature vector to be predicted, specifically:
Since the prediction of silicon rod growth rate needs to use the feature of future time period, the similar period filtered out is used
Data set predict the feature of future time period.
Feature vector to be predicted is made of 12 features in this programme:
Wherein, P is furnace pressure (MPa), TinFor intake air temperature (K), TsiIt is that silicon rod is straight for silicon rod surface temperature (K), D
Diameter (mm), V are intake velocity (m/s),For hydrogen inlet amount (m3),Trichlorosilane inlet amount (kg), TwallFor furnace
Inner wall temperature, N are silicon rod logarithm, Vsi1: the previous hour growth rate (mm/ of previous cycle synchronization in prediction period
h),Vsi2: growth rate (mm/h) of the previous cycle in prediction period in the same time, Vsi3: previous cycle is same in prediction period
The growth rate (mm/h) of the latter hour at one moment.
5. prediction result module
For the feature vector to be predicted to be input in prediction model, prediction result is exported.
Feature vector to be predicted obtains final prediction result through model, using 1h as a prediction point prediction future 12h in,
The growth rate of 12 points.
Claims (5)
1. silicon rod growth rate prediction model in a kind of polycrystalline silicon reducing furnace, which is characterized in that comprise the following modules:
Data acquisition and preprocessing module construct data for acquiring data relevant to silicon rod growth rate by sensor
Collection, and data cleansing is carried out to the data of acquisition；
Data set screening module, for being screened to data set of the data acquisition module after data acquisition, data cleansing,
Obtain training set and test set；
Trained and evaluation module is trained for promoting decision-tree model to gradient, utilizes average relative error and qualification
Rate assesses model, specifically includes:
(1) training set is inputted, sample number is M in training set；The number of iterations T and loss function L is set；The loss function is adopted
With square error loss function,Wherein y indicates that the true value of sample, y ' (x) indicate prediction
Value；
(2) weak learner is initialized
Find the constant value c for minimizing loss function:
In above formula, f0It (X) is weak learner, L () indicates loss function, yiIndicate the true output of i-th of sample；
(3) start iteration, in total iteration T times
(3-1) calculates negative gradient
The negative gradient of loss function when calculating each iteration: rti=yi-ft-1(Xi)；T=1,2...T, i=1,2...M, ft-1
(Xi) indicate the obtained strong learner of the t-1 times iteration to sample XiPredicted value；
(3-2) utilizes (Xi,rti) the weak regression tree of fitting, obtain the t weak regression tree；Its corresponding leaf node region is Rtj,
Middle j=1,2...J, leaf node number of the J for the t weak regression tree, t=1,2 ... T；
(3-3) is directed to leaf node region Rtj, find the best output valve c of fitting leaf nodetj:
(3-4) updates strong learner:
Wherein, ft-1(X) be the t-1 times iteration when strong learner, ftIt (X) is updated strong learner, the value range of v is
0 < v≤1, I expression indicator function, works as Xi∈RtjWhen, otherwise value 1 is 0；
(4) iteration terminates to obtain final strong learner:
Wherein, fT(X) the strong learner to be obtained after T iteration；
(5) strong learner is assessed using test set, obtains prediction model；
Module to be predicted, for inputting feature vector to be predicted；
Prediction result module exports prediction result for the feature vector to be predicted to be input in prediction model.
2. silicon rod growth rate prediction model in polycrystalline silicon reducing furnace as described in claim 1, which is characterized in that the acquisition
Related data, comprising: furnace pressure, intake velocity, the amounts of hydrogen of charging and the amount of trichlorosilane, silicon rod surface temperature, into
Temperature degree, the diameter of silicon rod, historical growth rate, silicon rod logarithm, furnace wall temperature；Using one group of these data as a sample
This.
3. silicon rod growth rate prediction model in polycrystalline silicon reducing furnace as described in claim 1, which is characterized in that pair
The data of acquisition carry out data cleansing, comprising:
Remember that the sample size acquired in total is m, for each of data set sample, if there are missing datas in sample, then
The missing data is replaced using the average value of corresponding data in previous production cycle, the latter production cycle, it is scarce to carry out
The processing of mistake value；For outlier processing, using statistical method, the exceptional value of every a kind of data is filtered out from data set, so
It handles using the exceptional value as missing values, or is corrected with the average value of every other same type data afterwards.
4. silicon rod growth rate prediction model in polycrystalline silicon reducing furnace as described in claim 1, which is characterized in that pair
Data acquisition module is acquired by data, the data set after data cleansing is screened, and obtains training set and test set, comprising:
The data in each sample concentrated to data are standardized, formula are as follows:
In above formula, xijIndicate the value of j-th of influence factor in i-th of sample, x 'ijIndicate xijAfter standardization
Value, max (xj)、min(xj) respectively indicate maximum value, the minimum value of all jth class influence factors in data set；Wherein, i=1,
2,…m；J=1,2 ... n, m indicate that sample size, n are the number of influence factor in sample；
Calculate the specific gravity p of j-th of influence factor in i-th of sampleij:
Calculate the entropy e of j-th of influence factorj:
In above formula,
Calculate the entropy weight w of j-th of influence factorj:
To obtain the weight vectors of each influence factor:
W=[w1 w2 w3...wn]
Calculate the initial matrix of each influence factor:
To initial matrix standardization processing:
Wherein:Wherein i=1,2 ... m, j=1,2 ... n；
If the feature vector of period to be predicted is y0=[y01 y02 ... y0j], then characteristics of time interval vector y to be predicted0With i-th
The feature vector y of sampleiIt is a in the degree of association of j-th of influence factorij
Wherein, y0jIndicate j-th of influence factor value of characteristics of time interval vector to be predicted, λ indicates resolution ratio；
Then degree of association Matrix C are as follows:
Wherein: amnIndicate the association angle value of n-th of influence factor of m-th of sample；
Weighted association judgment matrix are as follows:
Then projection value are as follows:
Wherein EiFor projection value of i-th of sample in the feature vector of period to be predicted；
According to projection value according to being arranged from big to small, threshold value is set, select all samples greater than threshold value as it is similar when
The data set of section, and training set and test set are divided on the basis of this data set.
5. silicon rod growth rate prediction model in polycrystalline silicon reducing furnace as described in claim 1, which is characterized in that the benefit
Strong learner is assessed with test set, obtains prediction model, comprising:
Calculate average relative error:
Wherein, K indicates the sum of the future position in predicted time section, yiIndicate that i moment silicon rod growth rate is true in predicted time section
Real value, y 'iIndicate i moment growth rate predicted value in predicted time section；
Qualification rate:
Wherein, Q (e < 4%) indicates the number of point of the relative error less than 4%, and K indicates future position sum；
If final strong learner meets the average relative error and qualification rate of setting, then finally prediction mould will be used as by strong learner
Type.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811454413.1A CN109543916B (en) | 2018-11-30 | 2018-11-30 | Model for estimating growth rate of silicon rod in polycrystalline silicon reduction furnace |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811454413.1A CN109543916B (en) | 2018-11-30 | 2018-11-30 | Model for estimating growth rate of silicon rod in polycrystalline silicon reduction furnace |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109543916A true CN109543916A (en) | 2019-03-29 |
| CN109543916B CN109543916B (en) | 2022-04-08 |
Family
ID=65851415
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811454413.1A Active CN109543916B (en) | 2018-11-30 | 2018-11-30 | Model for estimating growth rate of silicon rod in polycrystalline silicon reduction furnace |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109543916B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110321646A (en) * | 2019-07-10 | 2019-10-11 | 海默潘多拉数据科技(深圳)有限公司 | A kind of multiphase flow rates virtual metrology method promoting regression tree model based on gradient |
| CN110362558A (en) * | 2019-06-12 | 2019-10-22 | 广东工业大学 | A kind of energy consumption data cleaning method based on neighborhood propagation clustering |
| CN110470263A (en) * | 2019-08-02 | 2019-11-19 | 天津大学 | A kind of revolving body measurement system error compensation method based on gradient boosted tree |
| CN111784084A (en) * | 2020-08-17 | 2020-10-16 | 北京市城市规划设计研究院 | Travel generation prediction method, system and device based on gradient lifting decision tree |
| CN112097365A (en) * | 2020-07-10 | 2020-12-18 | 珠海派诺科技股份有限公司 | Air conditioner fault detection and identification method and device based on prediction and classification model |
| CN112985096A (en) * | 2021-03-05 | 2021-06-18 | 广州东兆信息科技有限公司 | Kiln ceramic quality monitoring system and method based on Internet of things cloud platform |
| CN113703411A (en) * | 2021-08-31 | 2021-11-26 | 亚洲硅业(青海)股份有限公司 | Polycrystalline silicon growth process monitoring system and method and polycrystalline silicon production system |
| CN116187112A (en) * | 2023-05-04 | 2023-05-30 | 北京大学 | Method and system for improving uniformity of thermal convection distribution of single crystal growth based on big data |
| CN117668740A (en) * | 2024-02-02 | 2024-03-08 | 浙江晶盛机电股份有限公司 | Sapphire long-speed abnormality detection method, device, electronic device and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102978687A (en) * | 2012-12-21 | 2013-03-20 | 英利集团有限公司 | Crystal growth method of polycrystalline silicon ingot |
| CN106521459A (en) * | 2016-08-17 | 2017-03-22 | 中山大学 | Optimization method of MOCVD equipment growth uniformity process parameters |
| US20180335538A1 (en) * | 2017-05-22 | 2018-11-22 | Schlumberger Technology Corporation | Resource Production Forecasting |
-
2018
- 2018-11-30 CN CN201811454413.1A patent/CN109543916B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102978687A (en) * | 2012-12-21 | 2013-03-20 | 英利集团有限公司 | Crystal growth method of polycrystalline silicon ingot |
| CN106521459A (en) * | 2016-08-17 | 2017-03-22 | 中山大学 | Optimization method of MOCVD equipment growth uniformity process parameters |
| US20180335538A1 (en) * | 2017-05-22 | 2018-11-22 | Schlumberger Technology Corporation | Resource Production Forecasting |
Non-Patent Citations (1)
| Title |
|---|
| 黄哲庆等: "一种新型多晶硅还原炉流动与传热的数值模拟", 《化工学报》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110362558B (en) * | 2019-06-12 | 2022-12-16 | 广东工业大学 | Energy consumption data cleaning method based on neighborhood propagation clustering |
| CN110362558A (en) * | 2019-06-12 | 2019-10-22 | 广东工业大学 | A kind of energy consumption data cleaning method based on neighborhood propagation clustering |
| CN110321646A (en) * | 2019-07-10 | 2019-10-11 | 海默潘多拉数据科技(深圳)有限公司 | A kind of multiphase flow rates virtual metrology method promoting regression tree model based on gradient |
| CN110470263A (en) * | 2019-08-02 | 2019-11-19 | 天津大学 | A kind of revolving body measurement system error compensation method based on gradient boosted tree |
| CN112097365A (en) * | 2020-07-10 | 2020-12-18 | 珠海派诺科技股份有限公司 | Air conditioner fault detection and identification method and device based on prediction and classification model |
| CN111784084A (en) * | 2020-08-17 | 2020-10-16 | 北京市城市规划设计研究院 | Travel generation prediction method, system and device based on gradient lifting decision tree |
| CN111784084B (en) * | 2020-08-17 | 2021-12-28 | 北京市城市规划设计研究院 | Travel generation prediction method, system and device based on gradient lifting decision tree |
| CN112985096B (en) * | 2021-03-05 | 2021-09-07 | 广州东兆信息科技有限公司 | Kiln ceramic quality monitoring system and method based on Internet of things cloud platform |
| CN112985096A (en) * | 2021-03-05 | 2021-06-18 | 广州东兆信息科技有限公司 | Kiln ceramic quality monitoring system and method based on Internet of things cloud platform |
| CN113703411A (en) * | 2021-08-31 | 2021-11-26 | 亚洲硅业(青海)股份有限公司 | Polycrystalline silicon growth process monitoring system and method and polycrystalline silicon production system |
| CN116187112A (en) * | 2023-05-04 | 2023-05-30 | 北京大学 | Method and system for improving uniformity of thermal convection distribution of single crystal growth based on big data |
| CN116187112B (en) * | 2023-05-04 | 2023-08-11 | 北京大学 | Method and system for improving uniformity of thermal convection distribution of single crystal growth based on big data |
| CN117668740A (en) * | 2024-02-02 | 2024-03-08 | 浙江晶盛机电股份有限公司 | Sapphire long-speed abnormality detection method, device, electronic device and storage medium |
| CN117668740B (en) * | 2024-02-02 | 2024-07-02 | 浙江晶盛机电股份有限公司 | Sapphire long-speed abnormality detection method, device, electronic device and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109543916B (en) | 2022-04-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109543916A (en) | Silicon rod growth rate prediction model in a kind of polycrystalline silicon reducing furnace | |
| CN109802430B (en) | Wind power grid control method based on LSTM-Attention network | |
| CN109147878B (en) | Soft measurement method for free calcium of cement clinker | |
| CN110378799A (en) | Aluminium oxide comprehensive production index decision-making technique based on multiple dimensioned depth convolutional network | |
| CN103345559B (en) | The dynamic evolutionary modeling method of aluminium electrolysis process art electrolysis cells energy consumption | |
| CN105574615B (en) | wavelet-BP neural network wind power prediction method based on spatial correlation and GA | |
| CN103514457B (en) | The multi-target improvement differential evolution model of Land use structure type when Ecology Restriction | |
| CN106503867A (en) | A kind of genetic algorithm least square wind power forecasting method | |
| CN110427654A (en) | A kind of predictive model of landslide construction method and system based on sensitiveness | |
| CN109669009A (en) | The adaptive extreme learning machine electronic nose drift rejection method in domain based on domain correction | |
| WO2023035727A1 (en) | Industrial process soft-measurement method based on federated incremental stochastic configuration network | |
| CN110347192B (en) | Glass furnace temperature intelligent prediction control method based on attention mechanism and self-encoder | |
| CN109284771A (en) | A kind of tomato growth model determination method and device | |
| CN110222714B (en) | Total solar irradiation resource prediction method based on ARMA and BP neural network | |
| CN110287540A (en) | A kind of photovoltaic cell parameter identification method based on elite masses' differential evolution algorithm | |
| CN112597430B (en) | Operation parameter optimization method for complex rectifying tower | |
| CN114119273A (en) | Park comprehensive energy system non-invasive load decomposition method and system | |
| CN110097929A (en) | A kind of blast furnace molten iron silicon content on-line prediction method | |
| CN114417740B (en) | Deep sea breeding situation sensing method | |
| CN116307039A (en) | Intelligent prediction method for photovoltaic output considering gas aberration anisotropy | |
| CN107729988A (en) | Blue-green alga bloom Forecasting Methodology based on dynamic depth confidence network | |
| Jin et al. | Soft sensor modeling for small data scenarios based on data enhancement and selective ensemble | |
| CN111950619B (en) | Active learning method based on dual-generation countermeasure network | |
| CN109472320B (en) | Method for automatically correcting variety parameters of crop growth period model under uncertain conditions | |
| CN110766234B (en) | Cement cooling process grate down pressure prediction method based on information fusion |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |