CN105160396B - A kind of method that neural network model is established using field data - Google Patents

A kind of method that neural network model is established using field data Download PDF

Info

Publication number
CN105160396B
CN105160396B CN201510392524.4A CN201510392524A CN105160396B CN 105160396 B CN105160396 B CN 105160396B CN 201510392524 A CN201510392524 A CN 201510392524A CN 105160396 B CN105160396 B CN 105160396B
Authority
CN
China
Prior art keywords
mrow
node
network
nodes
msub
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.)
Active
Application number
CN201510392524.4A
Other languages
Chinese (zh)
Other versions
CN105160396A (en
Inventor
雎刚
钱磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southeast University
Original Assignee
Southeast University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Southeast University filed Critical Southeast University
Priority to CN201510392524.4A priority Critical patent/CN105160396B/en
Publication of CN105160396A publication Critical patent/CN105160396A/en
Application granted granted Critical
Publication of CN105160396B publication Critical patent/CN105160396B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Landscapes

  • Complex Calculations (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of method that neural network model is established using field data, by using the substantial amounts of DCS data of scene operation, establishes neural network model, overcomes many drawbacks of traditional mechanisms modeling.In order to solve the problems, such as that neural network model generalization ability is poor, pruning algorithms are applied in RBF neural, the hidden node and input node to network carry out beta pruning, not only increase the generalization ability of network, while also determine the order of model.In order to avoid deleting input node by mistake, when carrying out beta pruning to input node, using the strategy of separated beta pruning, i.e., the process in input node is output and input and carry out beta pruning respectively, be thus avoided that the situation for all deleting the input of process in input node.

Description

A kind of method that neural network model is established using field data
Technical field
The present invention relates to the discrimination method in hot-working process control field, more particularly to thermal process.
Background technology
Thermal process have the characteristics that it is non-linear, when ductility, uncertainty and variable between relevance, it is difficult to establish essence True mathematical model, this brings very big difficulty to thermal control process.The model and field device that traditional modelling by mechanism obtains There is a certain distance in performance, because the performance of most equipment at scene is all non-linear, it is impossible to public by mathematics Formula obtains accurate model.Huge DCS system is assembled by the power plant of modernization, accumulates over a long period and acquires substantial amounts of operation number According to how from the rule between each variable of extracting data for containing bulk information, becoming current research hotspot.
RBF neural has the ability that can approach any nonlinear mapping, therefore is widely used in Nonlinear Modeling. But fixed formula is never had for the definite problem of neural network structure, and it can only be determined empirically, having no basis to follow, Its size is directly related to the generalization ability of neutral net.How the number of network node is reduced, and simplifying the structure of network becomes The hot spot of recent research.
The content of the invention
Goal of the invention：In order to overcome the deficiencies in the prior art, the present invention is provided one kind and is established using field data The method of neural network model.
Technical solution：In order to solve the above technical problems, one kind provided by the invention establishes neutral net using field data The method of model, comprises the following steps：
Step 1, live DCS data samplings；
Step 2, RBF neural identification；
The process in input node is output and input in step 2 and carries out beta pruning respectively.
Specifically, in the step 1, select some object under the same period or process first outputs and inputs DCS Data, input, output are respectively with variable u, y expressions, sampling period 5s；Then by variable u, y makees time delay processing respectively, respectively The numerical value at 1~n moment before each variable is taken, forms sample data：
Wherein, n be 3~5 natural number, N=3000.
Specifically, the RBF neural identification process of the step 2 is as follows：
1. selecting suitable network structure, suitable network parameter is chosen, including it is Hidden nodes HiddenUnitNum, hidden Node overlapping coefficient overlap, erasure signal signal=0；
2. using RBF algorithms training sample training network, when continuous m test error no longer declines, end is trained.Stop Only training condition is：
E (k) is current test error in formula, and E (k-i) is preceding test error several times, ε2For arbitrarily small positive reality Number, m take 5；
3. 6. if signal=0, turns, and otherwise, if the training error of current network has saltus step compared with previous network, I.e.：E-E ' ＞ δ
E is the training error of current network in formula, and E ' is the training error of previous network, and δ is arithmetic number, takes 5, then uses Previous network terminates as identification result, identification；Otherwise, turn 6.；
4. according to formula Calculate the sensitivity of hidden node, u (k-i) nodes and y (k-j) node；
Y in formula(p)For the network output of the P sample, d(p)For the target output of the P sample, whFor h-th of hidden node Connection weight, ohFor output of h-th of hidden node to the P sample, xkFor k input of the P sample, δhIt is hidden for h-th The corresponding width of node, h hidden node correspond to the value of k-th of input；
5. if the sensitivity for having minimum in hidden node, u (k-i) nodes and y (k-j) node meets formula
ρ in formulaminFor sensitivity minimum in hidden node or u (k-i) nodes or y (k-j) node, ρmFor hidden node or u (k-i) sensitivity of node or y (k-j) node, n are Hidden nodes or u (k-i) number of nodes or y (k-j) number of nodes, ε1To be small In 1 arithmetic number,
Corresponding minimum sensitivity node is then deleted, signal=1 is remembered, goes to 4.；Otherwise, beta pruning terminates, and identification terminates；⑥ Identification terminates；
Beneficial effect：The present invention establishes neural network model using the DCS data of scene operation, and by pruning algorithms application Into RBF neural, to input node using the strategy for separating beta pruning, i.e., the process in input node is output and input point Beta pruning is not carried out.Many drawbacks of traditional mechanisms modeling are overcome, solve the problems, such as that neural network model generalization ability is poor, it is right The hidden node and input node of network carry out beta pruning, not only increase the generalization ability of network, while also determine the rank of model It is secondary, avoid the situation for deleting input node by mistake.
Except the technical problem of invention described above solution, form the technical characteristic of technical solution and by these skills Caused by the technical characteristic of art scheme outside advantage, a kind of method that neural network model is established using field data of the invention The other technologies problem that can solve, the other technical characteristics included in technical solution and these technical characteristics bring excellent Point, will be described in more detail in conjunction with the embodiments.
Embodiment
Embodiment：
With reference to the live DCS data that certain power plant 600MW unit boiler reheating temperature is controlled device as an example, say Bright technical scheme implementation process is as follows：
Step 1：Live DCS data samplings；
1. selecting the DCS data of reheat steam temperature under the same period, sampling period 5s, the variable of selection includes：Reheating subtracts Temperature water spray valve opening μ, temperature T before reheating direct-contact desuperheater outlet guide1；
2. by variable u, T1Make time delay processing respectively, take the numerical value at 5 moment before each variable respectively, form sample number According to：
Step 2：RBF neural recognizes；
1. selecting suitable network structure, suitable network parameter, including Hidden nodes HiddenUnitNum=are chosen 15th, hidden node overlap coefficient overlap=1, erasure signal signal=0；
2. using RBF algorithms training sample training network, when continuous 5 test errors no longer decline, end is trained.Stop Only training condition is：
E (k) is current test error in formula, and E (k-i) is preceding test error several times, ε2=0.0001；
3. 6. if signal=0, turns, and otherwise, if the training error of current network has saltus step compared with previous network, I.e.：E-E ' ＞ δ
In formula E be current network training error, E ' be previous network training error, δ=5, then using previous network As identification result, identification terminates；Otherwise, turn 6.；
4. according to formula Calculate the sensitivity of hidden node, u (k-i) nodes and y (k-j) node；
Y in formula(p)For the network output of the P sample, d(p)For the target output of the P sample, whFor h-th of hidden node Connection weight, ohFor output of h-th of hidden node to the P sample, xkFor k input of the P sample, δhIt is hidden for h-th The corresponding width of node, h hidden node correspond to the value of k-th of input；
5. if the sensitivity for having minimum in hidden node, u (k-i) nodes and y (k-j) node meets formula
ρ in formulaminFor sensitivity minimum in hidden node or u (k-i) nodes or y (k-j) node, ρmFor hidden node or u (k-i) sensitivity of node or y (k-j) node, n are Hidden nodes or u (k-i) number of nodes or y (k-j) number of nodes, ε1= 0.1；
Corresponding minimum sensitivity node is then deleted, signal=1 is remembered, goes to 4.；Otherwise, beta pruning terminates, and identification terminates；⑥ Identification terminates；
Result before beta pruning and after beta pruning is compared, as a result such as following table：
Beta pruning Contrast on effect table
Network structure Training error Test error
Before beta pruning 10-15-1 0.3034 0.0854
After beta pruning 4-11-1 0.3201 0.0652
It can be drawn by the Contrast on effect of upper table, pruning algorithms proposed in this paper can simplify while precision is ensured The number of nodes of network.The input node of network is u (k-2), u (k-3), y (k-1) after beta pruning, y (k-2), the quantity of hidden node Substantially reduce, 11 are reduced to by original 15.

Claims (1)

  1. A kind of 1. method that neural network model is established using field data, it is characterised in that comprise the following steps：
    Step 1, live DCS data samplings；
    Select some object under the same period or process first outputs and inputs DCS data, and variable u is used in input, output respectively, Y expressions, sampling period 5s；Then by variable u, y makees time delay processing respectively, takes the number at 1~n moment before each variable respectively Value, forms sample data：
    <mrow> <mtable> <mtr> <mtd> <mrow> <mi>u</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>,</mo> <mi>u</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>-</mo> <mn>2</mn> <mo>)</mo> </mrow> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>u</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>-</mo> <mi>n</mi> <mo>)</mo> </mrow> <mo>,</mo> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <mi>y</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>,</mo> <mi>y</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>-</mo> <mn>2</mn> <mo>)</mo> </mrow> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>y</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>-</mo> <mi>n</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> </mtable> <mo>,</mo> <mi>y</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>,</mo> <mi>k</mi> <mo>=</mo> <mi>n</mi> <mo>+</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> <mo>+</mo> <mn>2</mn> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>N</mi> <mo>,</mo> </mrow>
    Wherein, n be 3~5 natural number, N=3000；
    Step 2, RBF neural identification
    RBF neural identification process is as follows：
    1. obtain modeling sample：WithFor input, y (k), k=n+1, n+2 ..., N are output；
    2. selecting suitable network structure, suitable network parameter, including Hidden nodes HiddenUnitNum, hidden node are chosen Overlap coefficient overlap, erasure signal signal=0；
    3. using RBF algorithms training sample training network, terminate to train when continuous m test error no longer declines, stop Training condition is：
    <mrow> <mo>|</mo> <mfrac> <mn>1</mn> <mi>m</mi> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mrow> <mi>m</mi> <mo>-</mo> <mn>1</mn> </mrow> </munderover> <mi>E</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>-</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>-</mo> <mi>E</mi> <mrow> <mo>(</mo> <mi>k</mi> <mo>)</mo> </mrow> <mo>|</mo> <mo>&lt;</mo> <msub> <mi>&amp;epsiv;</mi> <mn>2</mn> </msub> </mrow>
    E (k) is current test error in formula, and E (k-i) is preceding test error several times, ε2For arbitrarily small arithmetic number, m takes 5；
    4. if signal=0, turns 5., otherwise, if the training error of current network has saltus step compared with previous network, i.e.,： E-E'＞ δ
    E is the training error of current network in formula, and E' is the training error of previous network, and δ is arithmetic number, takes 5, then using previous Network terminates as identification result, identification；Otherwise, turn 5.；
    5. according to formula Meter Calculate the sensitivity of hidden node, u (k-i) nodes and y (k-j) node；
    Y in formula(p)For the network output of the P sample, d(p)For the target output of the P sample, whFor the company of h-th of hidden node Meet weights, ohFor output of h-th of hidden node to the P sample, xkFor k input of the P sample, δhFor h-th of hidden node Corresponding width, ch,kThe value of k-th of input is corresponded to for h-th of hidden node；
    6. if the sensitivity for having minimum in hidden node, u (k-i) nodes and y (k-j) node meets formula
    <mrow> <mfrac> <msub> <mi>&amp;rho;</mi> <mi>min</mi> </msub> <mrow> <mfrac> <mn>1</mn> <mi>n</mi> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <mo>|</mo> <msub> <mi>&amp;rho;</mi> <mi>m</mi> </msub> <mo>|</mo> </mrow> </mfrac> <mo>&lt;</mo> <msub> <mi>&amp;epsiv;</mi> <mn>1</mn> </msub> </mrow>
    ρ in formulaminFor sensitivity minimum in hidden node or u (k-i) nodes or y (k-j) node, ρmSaved for hidden node or u (k-i) The sensitivity of point or y (k-j) node, n are Hidden nodes or u (k-i) number of nodes or y (k-j) number of nodes, ε1For less than 1 just Real number,
    Corresponding minimum sensitivity node is then deleted, signal=1 is remembered, goes to 3.；Otherwise, beta pruning terminates, and identification terminates；
    7. identification terminates.
CN201510392524.4A 2015-07-06 2015-07-06 A kind of method that neural network model is established using field data Active CN105160396B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201510392524.4A CN105160396B (en) 2015-07-06 2015-07-06 A kind of method that neural network model is established using field data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201510392524.4A CN105160396B (en) 2015-07-06 2015-07-06 A kind of method that neural network model is established using field data

Publications (2)

Publication Number Publication Date
CN105160396A CN105160396A (en) 2015-12-16
CN105160396B true CN105160396B (en) 2018-04-24

Family

ID=54801248

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201510392524.4A Active CN105160396B (en) 2015-07-06 2015-07-06 A kind of method that neural network model is established using field data

Country Status (1)

Country Link
CN (1) CN105160396B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106054609B (en) * 2016-06-22 2019-04-23 东南大学 A kind of unit Stream temperature degree controlled process transfer function model discrimination method
CN106650928A (en) * 2016-10-11 2017-05-10 广州视源电子科技股份有限公司 Neural network optimization method and device
CN106548234A (en) * 2016-11-17 2017-03-29 北京图森互联科技有限责任公司 A kind of neural networks pruning method and device

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101017373A (en) * 2006-12-28 2007-08-15 浙江大学 Industrial process multiresolution softsensoring instrument and method thereof
KR101313822B1 (en) * 2012-08-07 2013-09-30 한국에너지기술연구원 Wind power density prediction method using neural network technique

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070150424A1 (en) * 2005-12-22 2007-06-28 Pegasus Technologies, Inc. Neural network model with clustering ensemble approach

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101017373A (en) * 2006-12-28 2007-08-15 浙江大学 Industrial process multiresolution softsensoring instrument and method thereof
KR101313822B1 (en) * 2012-08-07 2013-09-30 한국에너지기술연구원 Wind power density prediction method using neural network technique

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
"基于RBF神经网络的离心压缩机数据建模";张坤仁;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20150515;第C029-214页 *
"基于神经网络剪枝算法的热工过程模型辨识研究";尹婷婷 等;《江苏电机工程》;20080731;第11-13页 *
"基于神经网络的热工过程辨识";尹婷婷;《万方数据知识服务平台》;20101125;正文第13-26页 *

Also Published As

Publication number Publication date
CN105160396A (en) 2015-12-16

Similar Documents

Publication Publication Date Title
CN105160396B (en) A kind of method that neural network model is established using field data
CN103106535B (en) Method for solving collaborative filtering recommendation data sparsity based on neural network
CN104932274B (en) One kind coordinates control controlled device transfer function model discrimination method
CN107272412A (en) A kind of identifying approach of intermittent wind tunnel flow field control
CN108760592A (en) A kind of unburned carbon in flue dust On-line Measuring Method based on BP neural network
CN105184400A (en) Tobacco field soil moisture prediction method
CN105912822B (en) The neuron network simulation method of seawater desulfurization device desulfuration efficiency
CN101968832A (en) Coal ash fusion temperature forecasting method based on construction-pruning mixed optimizing RBF (Radial Basis Function) network
CN201476905U (en) Neural network PID temperature controlled thermocouple automatic verification system
CN105955014A (en) Method for controlling coke furnace chamber pressure based on distributed dynamic matrix control optimization
CN101887260A (en) Industrial process optimal control system and method for adaptive synchronization policy
CN103697958B (en) The real time measure method of coal unit drum outlet saturation steam mass rate
CN101529347A (en) Method and device for the identification of a delay-susceptible control path, control device, and computer program product
CN106292289B (en) The mixing elite Stochastic search optimization method of process industry control loop object
Park et al. Design of a nonlinear model-based controller with adaptive PI gains for robust control of a nuclear reactor
CN103793746A (en) Method for identifying parameters of coal-fired power plant boiler superheater model
CN105673094B (en) Turbine rotating speed control method based on active-disturbance-rejection control
CN107526294A (en) A kind of Nonlinear Delay dynamic system model INTELLIGENT IDENTIFICATION method
CN105607486A (en) General control method for series chemical integration time lag process
CN115907178A (en) Clean ecosystem CO 2 Method for predicting exchange amount
CN106950835A (en) Succinct Robust Two-Degree-of-Freedom proportional plus integral control method
CN114519308A (en) Method for determining river water and underground water interconversion lag response time influenced by river water and sand regulation
Opalka et al. Nonlinear state and unmeasured disturbance estimation for use in power plant superheaters control
CN109390946B (en) Optimal probability load flow rapid calculation method based on multi-parameter planning theory
CN110631003A (en) Reheated steam temperature adjusting method based on hierarchical scheduling multi-model predictive control

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant