CN106709197A

CN106709197A - Molten iron silicon content predicting method based on slide window T-S fuzzy neural network model

Info

Publication number: CN106709197A
Application number: CN201611269293.9A
Authority: CN
Inventors: 杨春节; 周恒�
Original assignee: Zhejiang University ZJU
Current assignee: Zhejiang University ZJU
Priority date: 2016-12-31
Filing date: 2016-12-31
Publication date: 2017-05-24

Abstract

The invention discloses a method for predicting the silicon content of molten iron based on a sliding window T-S fuzzy neural network model, which belongs to the field of industrial process monitoring, modeling and simulation. First, T-S fuzzy neural network is selected as the basic model of prediction; secondly, a sliding window model is added on the basis of this neural network, which can continuously update the training sample set to better track the change trend of silicon content; then According to actual experience and mutual information calculation, 11 parameters that have the greatest influence on the silicon content of molten iron are selected as the input of the model, and the silicon content of molten iron is taken as the output of the model; finally, the training samples are normalized and used to train the model, and the trained The model is used to predict the silicon content of molten iron during production. The time-varying, dynamic, non-linear, strong inertia and multi-scale characteristics of blast furnaces in the ironmaking process cause violent and unpredictable fluctuations in the silicon content of molten iron. Compared with the existing invention, the present invention has higher precision and smaller mean square error, and can be used for online real-time prediction.

Description

Molten iron silicon content Forecasting Methodology based on sliding window T-S fuzzy neural network models

Technical field

The invention belongs to industrial process monitoring, modeling and simulation field, more particularly to a kind of modified EMD-Elman nerves The method of neural network forecast molten iron silicon content.

Background technology

Complicated mass-and heat-transfer, heterogeneous reaction and seal in blast furnace so that blast furnace has complicated time-varying, dynamic, non- Linearly, strong inertia and multiple dimensioned characteristic, the ironmaking processes allowed in blast furnace turn into one of most complicated industrial processes.Blast furnace Interior HTHP, deep-etching, strongly disturbing environment so that we are difficult directly to measure the hot situation in stove.But, in molten iron Silicone content is linearly related to furnace temperature, can reflect the quality of molten iron, and people are generally represented with the size of silicone content in molten iron The height of furnace temperature.Silicone content is too high to represent that furnace temperature is too high, can consume extra fuel, and can reduce the yield of iron；And silicon contains The low expression in-furnace temperature of amount is relatively low, may trigger and freeze the accidents such as cylinder.Therefore, for the stable smooth operation of blast furnace, people are needed stove In certain zone of reasonableness, the prediction of silicone content is just particularly important temperature control system.

The change of silicon is main in blast furnace is made up of three below reaction：

1/2CO+O₂=1/2CO₂

SiO₂+ CO=SiO (g)+CO₂

SiO (g)+C=Si+CO

By Arrhenius equation, temperature and concentration have a great impact to chemical reaction rate, in the phase of silicon Close in reacting, it can be seen that the influence of temperature, oxygen concentration and carbonomonoxide concentration to molten iron silicon content is maximum.Therefore, have Scholar establishes the mechanism model of molten iron silicon content prediction by dynamics and thermodynamics, and they pay close attention to the heat in course of reaction Amount and the conservation of mass.But, because mass-and heat-transfer complicated in blast furnace, phase change and chemical reaction so that modelling by mechanism is little The content of silicon can accurately be predicted.

Nowadays, the development of detection means allows that we measure substantial amounts of data, and the fast development of computer technology makes Obtaining us can in a short time carry out substantial amounts of computing, and these technological progresses cause that the modeling based on data-driven becomes more Easily, the also modeling method as main flow.The model based on data-driven for having existed has neutral net, linear regression, mixes Ignorant and supporting vector machine model etc., they have oneself respective strong point in some aspects.For example, the chaos grain that Jiang is proposed Subgroup optimized algorithm can well predict the temperature in successive reaction kettle in pharmaceuticals industry.But, these models are all to set up It is determined that data set on, be not suitable for industrial on-line prediction.

T-S fuzzy neural networks possess very strong adaptive ability, can automatically update model structure parameter, and can correct The membership function of fuzzy subset, can well be used for the molten iron silicon content in PREDICTIVE CONTROL ironmaking processes.Slided by combining Dynamic window model, model can more new training sample set at any time, and then update the parameter and coefficient of T-S fuzzy neural networks.It is sliding The characteristics of dynamic window T-S fuzzy neural networks can be well adapted for dynamic, the non-linear and strong inertia of ironmaking processes, in molten iron Good performance is shown in the prediction of silicone content.

The content of the invention

For the weak point of existing Silicon Content Prediction in Process of Iron model, it is proposed that one kind is based on sliding window T-S fuzznets The molten iron silicon content Forecasting Methodology of network.The method is modeled from sliding window model and T-S fuzzy neural network models, and is chosen 11 major parameters as model input, using silicone content as model output.The method have hit rate higher and compared with Small mean square error, can provide accurately prediction for the operating personnel of blast furnace, help their advance operation blast furnaces, make blast furnace steady Determine direct motion.The method is comprised the steps of：

Step one：T-S fuzzy neural network models are chosen, and combines sliding window model, for the prediction of silicone content.

Step 2：The input that 11 parameters are chosen as model, silicone content conduct are calculated by practical experience and mutual information Output.

Step 3：After by model initialization, with normalized training sample training pattern, the model that will be trained is used for silicon The prediction of content.

The structure of the T-S fuzzy neural networks described in step one is as follows：

T-S fuzzy neural networks are constituted by four layers, are respectively input layer, obfuscation layer, fuzzy rule computation layer and output Layer.Wherein input is fuzzy, and exports what is be to determine, and this represents that output is the linear combination of input.T-S fuzzy neural networks It is defined as follows：

WhereinIt is fuzzy subset, y_iIt is the calculating output of fuzzy rule.

(1) obfuscation layer is that it is defined as follows based on probability density function μ：

X in formula_jIt is input variable,WithIt is center and the width of probability density function, k is the dimension of |input paramete, n It is the quantity of fuzzy subset.

(2) fuzzy rule computation layer is made up of following formula：

(3) output layer is calculated by following formula：

The learning algorithm of the T-S fuzzy neural networks described in step one is as follows：

(1) error calculation：

Wherein y_dIt is actual value, y_cIt is predicted value, e is both differences.

(2) coefficient amendment：

In formulaIt is the coefficient of T-S fuzzy neural networks, and α is its learning rate.

(3) parameters revision：

Sliding window model principle described in step one is as follows：

Sliding window model is built upon in a kind of hypothesis, i.e., current output depends on current input, and is input into defeated Mapping ruler between going out can be obtained by historical data.According to this it is assumed that we preset a certain amount of training set Sample, is then continuously updated sample data and gives up earliest data point.With the slip of window, T-S fuzzy neural networks Its structural parameters can be constantly updated and newest predicted value is given.

The selection process of the input variable described in step 2 is as follows：

Mutual information is a kind of important method of test variable correlation, and Kraskov proposes a kind of k-NN methods can be very It is convenient to be used for calculating mutual information, comprise the following steps that described：

K is the number of neighbour given at the beginning in formula, and ψ is that Digamma functions can be expressed as：

ψ (x)=Γ (x)^-1dΓ(x)/dx

It obeys following iterative relation：

ψ (x+1)=ψ (x)+1/x

Ψ (1)=- C, C=0.5772156...

In order to obtain n_xAnd n_y, it is necessary to calculate sample z_iAnd z_jThe distance between d_i,j：

d_i,j=| | z_i-z_j||:d_i,j1≤d_i,j2≤d_i,j3...

||z_i-z_j| |=max | | x_i-x_j||,||y_i-y_j||}

As ε (i)=max { ε_x(i),ε_y(i) }, ε (i)/2 are taken as z_iWith the distance of k rank neighbours.Obviously, n_xI () is to x_i Distance is less than the number of the point of ε (i)/2, n_yI () is to y_iNumber of the distance less than the point of ε (i)/2.

Advise that we have chosen input of 11 variables as model by the practical experience of execute-in-place engineer, it Be respectively top pressure, top temperature, gas permeability, coal powder injection, oxygen enrichment percentage, full tower pressure difference, hot-blast pressure, hot blast temperature, hot air flow, Air humidity and previous stove silicone content.

Method for normalizing described in step 3 is as follows：

The present invention has advantages below：

1st, for the time-varying of blast furnace, dynamic, non-linear, strong inertia and multiple dimensioned characteristic in ironmaking processes, tool has been selected There are the T-S fuzzy neural networks of very strongly-adaptive, it has very strong learning ability, can find out latent between input and output In contact.Additionally, by adding sliding window, model can well track the variation tendency of molten iron silicon content, improve pre- The precision of survey.

2nd, calculated by operating experience and mutual information, selected top pressure, top temperature, gas permeability, coal powder injection, oxygen enrichment percentage, complete Tower pressure difference, hot-blast pressure, hot blast temperature, hot air flow, air humidity and previous stove silicone content etc. influence most on current silicone content 11 big parameters can make full use of the respective advantage of modelling by mechanism and data-driven modeling as the input variable of model.

Brief description of the drawings

Fig. 1 is the structural representation of T-S fuzzy neural networks,

Fig. 2 is the schematic diagram of sliding window,

Fig. 3 is the molten iron silicon content of 1000 stoves,

Fig. 4 is that this method predicts the outcome to molten iron silicon content.

Specific embodiment

The present invention proposes a kind of molten iron silicon content Forecasting Methodology based on sliding window T-S fuzzy neural networks, the party Method is comprised the steps of：

Step one：T-S fuzzy neural network models (such as Fig. 1) is chosen, and combines sliding window model (such as Fig. 2), be used for The prediction of silicone content.

Step 3：After by model initialization, with normalized training sample training pattern, the model that will be trained is used for silicon The prediction (such as Fig. 4) of content.

WhereinIt is fuzzy subset, y_iIt is the calculating output of fuzzy rule.

(2) fuzzy rule computation layer is made up of following formula：

(3) output layer is calculated by following formula：

(1) error calculation：

Wherein y_dIt is actual value, y_cIt is predicted value, e is both differences.

(2) coefficient amendment：

(3) parameters revision：

Sliding window model principle described in step one is as follows：

ψ (x)=Γ (x)^-1dΓ(x)/dx

It obeys following iterative relation：

ψ (x+1)=ψ (x)+1/x

Ψ (1)=- C, C=0.5772156...

d_i,j=| | z_i-z_j||:d_i,j1≤d_i,j2≤d_i,j3...

||z_i-z_j| |=max | | x_i-x_j||,||y_i-y_j||}

Method for normalizing described in step 3 is as follows：

Embodiment

In steel manufacture process, blast furnace ironmaking is all good link of crucial importance, and the energy consumption of its consumption accounts for whole flow process 70%, thus blast furnace stable smooth operation be whole production process safe and highly efficient operation guarantee.Because the environment in blast furnace is extremely Badly, HTHP deep-etching so that conventional measurement means are difficult to carry out, and operating personnel are difficult to know the actual heat in blast furnace Situation, when tapping a blast furnace, molten iron loss amount of heat can not react actual furnace temperature.People are generally using molten iron silicon content come reacting furnace Interior actual state, thus the prediction of molten iron silicon content just seems of crucial importance, accurately prediction not only assists in operating personnel Rational regulation operating parameter, moreover it is possible to instruct blast furnace stable smooth operation.

We verify the accuracy of the model for proposing by studying No. 2 the 1000 of blast furnace groups of data (shown in Fig. 3) of Liu Gang. Below, we are explained in detail with reference to detailed process to implementation steps：

Advise that we have chosen input of 11 variables as model by the practical experience of execute-in-place engineer, it Be respectively top pressure, top temperature, gas permeability, coal powder injection, oxygen enrichment percentage, full tower pressure difference, hot-blast pressure, hot blast temperature, hot air flow, Air humidity and previous stove silicone content.Their association relationships with current silicone content are calculated, following result is obtained：

Numbering	Variable	Unit	Mutual information
				1	Top pressure	kPa	0.12
2	Top temperature	℃	0.22
				3	Gas permeability	m³/min·kPa	0.14
4	Coal powder injection	t/h	0.29
				5	Oxygen enrichment percentage	Vol%	0.21
6	Full tower pressure difference	kPa	0.10
				7	Hot-blast pressure	kPa	0.15
8	Hot blast temperature	℃	0.32
				9	Hot air flow	m³/min	0.13
10	Air humidity	Vol%	0.08
				11	Previous stove silicone content	Wt%	0.45

Association relationship is between 0 to 1, and two correlation of variables of bigger expression are stronger.Previous stove silicon contains as can be seen from the table Amount contacts maximum with current silicone content, but the influence of remaining variable can not be ignored.

Method for normalizing described in step 3 is as follows：

The sample size that we set training data is 400, and test data set is 50.With prediction hit rate J and mean square error MSE two indices verify the precision of model prediction：

In actual production process, predicated error can meet requirement less than 0.1.It is proposed that the hit rate of model reach To 90%, mean square error is 0.0023.With precision very high, it is entirely capable of meeting the demand of actual production.

Above-described embodiment is used for illustrating the present invention, rather than limiting the invention, in spirit of the invention and In scope of the claims, any modifications and changes made to the present invention belong to protection scope of the present invention.

Claims

1. a method for predicting molten iron silicon content based on sliding window T-S fuzzy neural network, is characterized in that, step is as follows:

Step 1: Select the T-S fuzzy neural network model and combine the sliding window model for the prediction of silicon content;

Step 2: Through actual experience and mutual information calculation, select 11 parameters as the input of the model, and the silicon content as the output. The parameters mentioned are top pressure, furnace top temperature, air permeability, coal injection, oxygen enrichment rate, and total tower pressure difference, hot air pressure, hot air temperature, hot air flow, air humidity and silicon content of the previous furnace;

Step 3: After initializing the model, train the model with normalized training samples, and use the trained model to predict the silicon content;

The structure of the T-S fuzzy neural network described in step one is as follows:

The T-S fuzzy neural network consists of four layers, namely the input layer, the fuzzy layer, the fuzzy rule calculation layer and the output layer, where the input is fuzzy and the output is definite, which means that the output is a linear combination of the input, T-S fuzzy neural network A network is defined as follows:

{R R}^{i i} : : I I f f {x x}_{11} i i s the s {A A}_{11}^{i i},, ... ...,, {x x}_{k k} i i s the s {A A}_{k k}^{i i},, t t h h e e n no {y the y}_{i i} = = {p p}_{00}^{i i} + + {p p}_{11}^{i i} {x x}_{11} + + ... ... + + {p p}_{k k}^{i i} {x x}_{k k}

in is a fuzzy subset, and y _i is the calculation output of fuzzy rules;

(1) The fuzzy layer is based on the probability density function μ, which is defined as follows:

{μ μ}_{{A A}_{j j}^{i i}} = = {e e}^{- - {(({x x}_{j j} - - {c c}_{j j}^{i i}))}^{22} / / {b b}_{j j}^{i i}},, ((i i = = 11,, ... ...,, n no;; j j = = 11,, ... ...,, k k))

where x _j is the input variable, with is the center and width of the probability density function, k is the dimension of the input parameter, n is the number of fuzzy subsets;

(2) The fuzzy rule calculation layer consists of the following formula:

{ω ω}^{i i} = = {μ μ}_{{A A}_{j j}^{11}} (({x x}_{11})) \times \times {μ μ}_{{A A}_{j j}^{22}} (({x x}_{22})) \times \times ... ... \times \times {μ μ}_{{A A}_{j j}^{k k}} (({x x}_{k k})),, ((i i = = 11,, ... ...,, n no))

(3) The output layer is calculated by the following formula:

{y the y}_{i i} = = \frac{{Σ Σ}_{i i = = 11}^{n no} {ω ω}^{i i} {y the y}_{i i}}{{Σ Σ}_{i i = = 11}^{n no} {ω ω}^{i i}} . .

2. method according to claim 1, is characterized in that, the learning algorithm of the T-S fuzzy neural network described in step one is as follows:

⑴ Error calculation:

e e = = \frac{11}{22} {(({y the y}_{d d} - - {y the y}_{c c}))}^{22}

Where y _d is the actual value, y _c is the predicted value, and e is the difference between the two;

⑵ Coefficient correction:

{p p}_{j j}^{i i} = = {p p}_{j j}^{i i} ((k k - - 11)) - - α α \frac{\partial \partial e e}{\partial \partial {p p}_{j j}^{i i}}

\frac{\partial \partial e e}{\partial \partial {p p}_{j j}^{i i}} = = \frac{(({y the y}_{d d} - - {y the y}_{c c})) {ω ω}^{i i}}{{Σ Σ}_{i i = = 11}^{m m} {ω ω}^{i i} \cdot &Center Dot; {x x}_{j j}}

In the formula is the coefficient of TS fuzzy neural network, and α is its learning rate;

⑶ Parameter correction:

\begin{matrix} {c c}_{j j}^{i i} ((k k)) = = {c c}_{j j}^{i i} ((k k - - 11)) - - β β \frac{\partial \partial e e}{\partial \partial {c c}_{j j}^{i i}} \\ {b b}_{j j}^{i i} ((k k)) = = {b b}_{j j}^{i i} ((k k - - 11)) - - β β \frac{\partial \partial e e}{\partial \partial {b b}_{j j}^{i i}} \end{matrix} . .

3. The method according to claim 1, wherein the principle of the sliding window model described in step 1 is as follows:

The sliding window model is based on the assumption that the current output depends on the current input, and the mapping rules between input and output can be obtained from historical data. According to this assumption, a certain amount of training set samples are preset, and then The sample data is constantly updated and the earliest data points are discarded. As the window slides, the T-S fuzzy neural network will constantly update its structural parameters and give the latest prediction value.

4. method according to claim 1, is characterized in that, the selection process of the input variable described in step 2 is as follows:

Mutual information is a method to test the correlation of variables, and the specific steps are as follows:

\begin{matrix} I I ((X x;; Y Y)) \\ = = \underset{x x &Element; &Element; X x}{Σ Σ} \underset{y the y &Element; &Element; Y Y}{Σ Σ} p p ((x x,, y the y)) {log log}_{b b} \frac{p p ((x x,, y the y))}{p p ((x x)) p p ((y the y))} \\ = = ψ ψ ((k k)) - - < < ψ ψ (({n no}_{x x} + + 11)) + + ψ ψ (({n no}_{y the y} + + 11)) > > + + ψ ψ ((N N)) \end{matrix}

In the formula, k is the number of neighbors given at the beginning, and ψ is the Digamma function, which can be expressed as:

ψ(x)=Γ(x) ^-1 dΓ(x)/dx

It obeys the following iteration relation:

ψ(x+1)=ψ(x)+1/x

Ψ(1)=-C,C=0.5772156...

< < ... ... > > = = {N N}^{- - 11} {Σ Σ}_{i i = = 11}^{N N} E E. [[... ... ((i i))]]

In order to get n _x and n _y , it is necessary to calculate the distance d _i,j between samples z _i and z _j :

d _i,j ＝||z _i -z _j ||:d _i,j1 ≤d _i,j2 ≤d _i,j3 …

||z _i -z _j ||＝max{||x _i -x _j ||,||y _i -y _j ||}

When ε(i)=max{ε _x (i),ε _y (i)}, ε(i)/2 is regarded as the distance between z _i and k-order neighbors, and n _x (i) is the distance to x _i less than The number of points of ε(i)/2, n _y (i) is the number of points whose distance to y _i is less than ε(i)/2;

According to the actual experience of the field operation engineer, 11 variables were selected as the input of the model, and the mutual information value between them and the current silicon content was calculated. The mutual information value ranged from 0 to 1. The larger the value, the stronger the correlation between the two variables.

5. method according to claim 1, is characterized in that, the normalization method described in step 3 is as follows:

y the y = = \frac{(({y the y}_{max max} - - {y the y}_{min min})) ((x x - - {x x}_{min min}))}{{x x}_{max max} - - {x x}_{min min}} + + {y the y}_{min min},, (({y the y}_{m m i i n no} = = - - 11,, {y the y}_{m m a a x x} = = 11)) . .

6. The method according to claim 1, characterized in that the model is suitable for the time-varying, dynamic, nonlinear, strong inertial and multi-scale characteristics of the blast furnace in the ironmaking process.