CN103679268A - Blast furnace slag viscosity prediction method - Google Patents

Abstract

The invention discloses a blast furnace slag viscosity prediction method which comprises the following steps: a neural network construction step, wherein a constructed neural network structure comprises an input layer, a hidden layer and an output layer, which are connected through connecting weights, and the hidden layer has a threshold value; an initialization step of acquiring training data, and initializing the training data, the connecting weights and the threshold value; a neural network self-learning step of sequentially extracting the training data, regulating the connecting weights and the threshold value according to an error and a local gradient until all the training data is used, and storing final values of the connecting weights and the threshold value; a furnace slag viscosity prediction step of fixing the connecting weights and the threshold value of the neural network structure at the final values, inputting the relative content of each component in actual furnace slag, and outputting predicted furnace slag viscosity at different temperature, melting temperature and a desulphurization coefficient. According to the method, the error can be stabilized to be about +/-2 percent, and is remarkably improved compared with that of +/-6 percent of a conventional mode.

Description

BF Slag Viscosity forecasting procedure

Technical field

The present invention relates to field of metallurgy, relate in particular to a kind of BF Slag Viscosity forecasting procedure.

Background technology

Existing blast furnace slag viscosity forecasting procedure is mainly experimental formula method, or be called regression analysis, by existing blast furnace slag viscosity data is carried out to linear regression analysis, show that slag viscosity and melting temperature are about the linear formula of slag composition, it is experimental formula, then by experimental formula, slag is carried out to performance prediction, the method simple operation, easily grasps.Shortcoming is that error is larger, and the composition range of application is narrower, and adaptability is poor.

For example, the Chinese patent that the patent No. is ZL200820100327.6 has disclosed a kind of high temperature viscometer of measuring slag and molten steel viscosity.This viscosity meter utilizes rotating cylindrical body (molybdenum probe) method, and the moment of rotating in slag by measuring sonde is calculated the viscosity number of slag indirectly.This patent only provides a kind of method of viscosity measurement, does not have the function of before slag viscosity is measured, slag viscosity being carried out to performance prediction.

Again such as, the Chinese patent application that publication number is CN102479290A has disclosed a kind of method of calculating the melting temperature of slag.The method is utilized the method for mathematical statistics, after obtaining many groups viscosity of slag and the data of institute's corresponding temperature thereof, obtain the data of many group viscosity and institute's corresponding temperature thereof and the funtcional relationship that viscosity with temperature changes, simulate temperature-viscosity curve, the coordinate system being arranged at this temperature-viscosity curve, and get the melting temperature that slope is described slag for the corresponding temperature of-1/ straight line of (50～70) and the point of contact of temperature-viscosity curve.The method utilization be the nonlinear relationship that mathematical statistics method analysis draws viscosity and temperature, error is larger, is used in slag viscosity forecast, can not adapt to the variation of slag composition, is a kind of Forecasting Methodology that cannot self.

Techniques of Neural Network research originates in the phase at the end of the eighties in last century, and theory is ripe so far, also has a lot of application examples, and has obtained good application achievements.Such as the Multi-layered Feedforward Networks based on BP algorithm, be used in fields such as catalyst formulation modeling, coal and gas outbursts Prediction, coal pulverizer monitoring material position, Project Investment Risk Evaluation, atmosphere quality evaluation.

But, up to the present, yet there are no both at home and abroad neural network model be used in to blast-furnace slag performance prediction field.In actual blast furnace ironmaking process, if increase lump ore ratio, can make Al in slag ₂o ₃content raises, the quantity of slag increases, slag fluidity variation.Now, traditional Performance of Slag forecast based on experimental formula cannot be reacted this variation, and forecast result there will be larger error.Therefore need to a kind ofly can promptly and accurately predict the parameters such as slag viscosity and melting degree warm in nature, for blast furnace process provides rational slag system to form and the Performance of Slag forecasting model of slagging regime, for improving blast furnace lump ore ratio, reduce the quantity of slag and reducing raw materials cost, provide foundation.

Summary of the invention

For the problems referred to above, the object of the invention is to forecast in time Performance of Slag, provide reasonable slag system to form and slagging process.The present invention proposes a kind of BF Slag Viscosity forecasting procedure of setting up based on improving the neural network model of BP algorithm.

According to one embodiment of the invention, a kind of BF Slag Viscosity forecasting procedure is proposed, comprise following step:

Neuroid construction step, the neuroid structure building comprises input layer, hidden layer and output layer, between input layer, hidden layer and output layer, by connecting weights, connect, hidden layer has threshold value, the input parameter of input layer is the percentage composition of each composition in slag, and the output parameter of output layer is slag viscosity, melting temperature and the desulfurization coefficient under different temperatures;

Initialization step, obtains training data, and to training data, connect weights and threshold value is carried out initialization, described training data comprises reference input parameter and with reference to output parameter;

Neuroid self study step, extract a training data, reference input parameter is wherein imported to neuroid, via input layer, hidden layer and output layer output training output parameter, relatively train the reference output parameter in output parameter and this training data, calculate error and the partial gradient of each layer, according to error, to set step-length adjustment, be connected weights and threshold value with partial gradient, according to partial gradient adjustment, set step-length, extract next training data and repeat said process, until all training datas are all used, preserve the end value that connects weights and threshold value,

Slag viscosity forecast step, is fixed on end value by connection weights and the threshold value of neuroid, inputs the percentage composition of each composition in actual slag, slag viscosity, melting temperature and desulfurization coefficient under the different temperatures of output forecast.

In one embodiment, in slag, the percentage composition of each composition comprises SiO ₂, Al ₂o ₃, CaO, MgO percentage composition; Slag viscosity under different temperatures comprises the slag viscosity η of 1500 ℃ _{1500 ℃}, the slag viscosity η of 1475 ℃ _{1475 ℃}slag viscosity η with 1450 ℃ _{1450 ℃}.

In one embodiment, training data is carried out to initialization and comprise training data is normalized, to connecting weights, carry out initialization with threshold value and comprise that by being connected weights and Threshold be initial value.

In one embodiment, according to partial gradient adjustment, set step-length and comprise: to partial gradient subsequent iteration twice, if two sub-symbols are identical, increase and set step-length; If two sub-symbols are contrary, reduce to set step-length.

The BF Slag Viscosity forecasting procedure that the present invention proposes realizes based on neuroid, can be stabilized in ± 2% left and right of error, than traditional error ± 6% based on linear regression formula mode, be significantly improved, neuroid has self-learning capability simultaneously, can use the scope of forecasting widely.

Accompanying drawing explanation

Fig. 1 has disclosed according to the process flow diagram of the BF Slag Viscosity forecasting procedure of one embodiment of the invention.

Fig. 2 has disclosed according to the input and output parameter of the BF Slag Viscosity forecasting procedure of a specific implementation of the present invention.

Fig. 3 has disclosed according to the model learning flow process based on improving BP algorithm in the BF Slag Viscosity forecasting procedure of a specific implementation of the present invention.

Fig. 4 has disclosed the error statistics that adopts BF Slag Viscosity forecasting procedure of the present invention.

Fig. 5 has disclosed the error statistics that adopts traditional linear regression formula mode to forecast.

Embodiment

The performance of accurate forecast blast-furnace slag, as viscosity and melting temperature etc., the composition of molten iron and quality, the usage factor of blast furnace, coke ratio coal are compared etc. to blast furnace technical indicator very large improvement effect.It is little that neuroid has prediction error, the features such as ability of strong adaptability and self-teaching, while applying to blast furnace slag performance prediction, can pass through after the study of Performance of Slag database, according to slag composition, the parameters such as the viscosity of accurate forecast slag discharging and melting temperature, the feature that simultaneously can also constantly change according to slag composition, autonomous learning, upgrade the corresponding relation (renewal connection weight) of output layer (viscosity of slag and melting temperature etc.) and input layer (slag composition), improve the accuracy of forecast, in error can be narrowed down to ± 2% scope, meet the precision needs of blast furnace process completely.

Performance of Slag forecasting model system based on artificial neural network is compared with traditional experimental formula method, has following characteristics:

Prediction error is little: model, in the process of learning sample, by continuous adjustment weights, makes the function of model approach gradually target, error is reduced in allowed band, therefore,, as long as there are enough learning times, the data error of this model prediction can reach accuracy requirement substantially.And the resultant error of traditional method forecast is larger, reason is that the linear relationship of input and output is poor, utilizes result and the actual value that formula calculates to depart from larger.

Strong adaptability: the scope of application of neural network model is larger, reason is that model passes through sample learning, adjusts weights, can adapt to the input numerical value that fluctuation is larger.And traditional experimental formula is the regression function obtaining in certain input range, exceed the scope of application, forecast result has larger error.

There is ability of self-teaching: after learning parameter (hidden layer number, learning time, forecast precision) is set, model can be according to the variation of sample database and autonomous learning, and the result after study is preserved, with this, guarantee that model can adapt to the variation of data all the time, improve forecast precision.

The principle of neural network model is as follows:

Neural network model is the nonlinear function model of input more than, single output, is divided into three-decker: input layer, hidden layer, output layer, every one deck directly connects by weights.It has two with by classic method, carry out the diverse character of information processing: 1. neuroid is self-adaptation and can being trained, and it has self-adjusting is self-learning capability.If last output is incorrect, system can be adjusted weights and be added to each and input up to produce a new result, so repeatedly, until reach desired output.2. neuroid structure itself has just determined that it is large-scale parallel mechanism.Because it is data-driven, therefore its processing speed is faster than classic method.

Neural network model can be used following function representation:

$c_j=f\left(\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{b}_i\right)=f_1[\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}\·f_2\left(\underset{h=1}{\overset{l}{\mathrm{\Σ}}}{v}_{\mathrm{hi}}{a}_h\right)]$

$b_i=f_2\left(\underset{h=1}{\overset{l}{\mathrm{\Σ}}}{v}_{\mathrm{hi}}{a}_h\right)$

C in formula _joutput for model;

B _ifor hidden layer input item;

A _hfor input layer input.

The present invention is intended to propose a kind of BF Slag Viscosity forecasting procedure based on neural network model, and Fig. 1 has disclosed according to the process flow diagram of the BF Slag Viscosity forecasting procedure of one embodiment of the invention.The method comprises following step:

102. neuroid construction steps, the neuroid structure building comprises input layer, hidden layer and output layer, between input layer, hidden layer and output layer, by connecting weights, connect, hidden layer has threshold value, the input parameter of input layer is the percentage composition of each composition in slag, and the output parameter of output layer is slag viscosity, melting temperature and the desulfurization coefficient under different temperatures.In one embodiment, in slag, the percentage composition of each composition comprises SiO ₂, Al ₂o ₃, CaO, MgO percentage composition.Slag viscosity under different temperatures comprises the slag viscosity η of 1500 ℃ _{1500 ℃}, the slag viscosity η of 1475 ℃ _{1475 ℃}slag viscosity η with 1450 ℃ _{1450 ℃}.

104. initialization steps, obtain training data, and to training data, connect weights and threshold value is carried out initialization, training data comprises reference input parameter and with reference to output parameter.In one embodiment,, training data is carried out to initialization and comprise training data is normalized, to connecting weights, carry out initialization with threshold value and comprise that by being connected weights and Threshold be initial value.

106. neuroid self study steps, extract a training data, reference input parameter is wherein imported to neuroid, via input layer, hidden layer and output layer output training output parameter, relatively train the reference output parameter in output parameter and this training data, calculate error and the partial gradient of each layer, according to error, to set step-length adjustment, be connected weights and threshold value with partial gradient, according to partial gradient adjustment, set step-length, extract next training data and repeat said process, until all training datas are all used, preserve the end value that connects weights and threshold value.In one embodiment, to partial gradient subsequent iteration twice, if two sub-symbols are identical, increase and set step-length; If two sub-symbols are contrary, reduce to set step-length.

108. slag viscosity forecast step, is fixed on end value by connection weights and the threshold value of neuroid, inputs the percentage composition of each composition in actual slag, slag viscosity, melting temperature and desulfurization coefficient under the different temperatures of output forecast.

Introduce according to the BF Slag Viscosity forecasting procedure of a specific implementation of the present invention below.This BF Slag Viscosity forecasting procedure is the algorithm based on utilizing the Performance of Slag forecast system of BP algorithm.

The BP algorithm of standard: error Back-Propagation (Error Back Propagation, be called for short EBP) or claim contrary (the Back Fropagation of propagation, be called for short BP), the error that network output is occurred is summed up as the mistake of each connection weight, thereby by the error of output layer unit is successively obtained to the reference error of each layer of unit to adjust corresponding connection weight to the reverse propagation of defeated people's layer to share to each layer of unit.

The most basic BP network is three layers of feedforward, i.e. forward connection between input layer, hidden layer and output layer unit.Conventionally, network can have a plurality of hidden layers, and network, by multilayer error correction gradient descent method off-line learning, moves by discrete time mode.BP network application is forecast to field at BF Slag Viscosity, and the input parameter of input layer is the constituent of slag, and the output parameter of output layer is slag viscosity, melting temperature Tm and the desulfurization coefficient Ls under different temperatures.Hidden layer comprises that threshold value θ and model calculate.Fig. 2 has disclosed according to the input and output parameter of the BF Slag Viscosity forecasting procedure of a specific implementation of the present invention.

Error Back-Propagation study completes from being input to the mapping of output cost function minimization process by one.Conventionally, cost function is defined as the error sum of squares that on all input patterns, output layer unit is wished output and actual output.

Error Back-Propagation study can be divided into two stages.In first stage, the network input for given, by its forward-propagating, obtains the actual output (activation value) of unit by existing connection weight; In second stage, first calculate the vague generalization error of each unit of output layer, these errors are successively propagated to input layer direction, to obtain, adjust each required elements reference error of each connection weight.

The derivation of error Back-Propagation learning algorithm below.

If E _kfor supplying a pattern to (A to network _k, C _k) time cost function on output layer, the global cost function on whole pattern drill collection is:

$E=\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{E}_k---(1-1)$

For k pattern pair, output layer unit j is weighted to:

$c_{\mathrm{jnet}}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{b}_i---(1-2)$

The actual of this unit is output as:

c _j=f(c _jnet) （1-3）

The weighting of face hidden layer unit i is input as:

$b_{\mathrm{inet}}=\underset{h=1}{\overset{L}{\mathrm{\Σ}}}{v}_{\mathrm{hi}}{a}_h---(1-4)$

The actual of this unit is output as:

b _i=f(b _inet) （1-5）

Function in formula (1-3) and formula (1-5) is differentiable non-decreasing function.

For output unit j, definition vague generalization error is:

$d_j=-\frac{{\∂E}_k}{{\∂c}_{\mathrm{jnet}}}---(1-6)$

Formula (1-6) can be write as form below:

$d_j=-\frac{{\∂E}_k}{{\∂c}_j}\frac{{\∂c}_j}{{\∂c}_{\mathrm{jnet}}}=-\frac{{\∂E}_k}{{\∂c}_j}{f}^{\mathrm{\γ}}\left(c_{\mathrm{inet}}\right)---(1-7)$

For implicit unit i, define equally vague generalization error and be:

$e_i=\frac{{\∂E}_k}{{\∂b}_{\mathrm{inet}}}---(1-8)$

Similarly, e _ithere is form below:

$e_i=-\frac{{\∂E}_k}{{\∂b}_i}\frac{{\∂b}_i}{{\∂b}_{\mathrm{inet}}}$

$=f^{\′}\left(b_{\mathrm{inet}}\right)(-\frac{{\∂E}_k}{{\∂b}_i})$

$=f^{\′}\left(b_{\mathrm{inet}}\right)(-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{{\∂E}_k}{{\∂c}_{\mathrm{jnet}}}\frac{{\∂c}_{\mathrm{jnet}}}{{\∂b}_i})$

$=f^{\′}\left(b_{\mathrm{inet}}\right)\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{d}_j{w}_{\mathrm{ij}}---(1-9)$

Formula (1-9) can be considered the contrary error that propagates into this layer of unit of anterior layer elemental error.

At existing connection weight w _ijand v _hiunder, in order to reduce cost function E _k, need to determine how to change connection weight.This can be completed by Gradient Descent rule (even if the variation of connection weight is proportional to negative gradient).Therefore, have:

${\mathrm{\Δw}}_{\mathrm{ij}}=-\mathrm{\α}\frac{{\∂E}_k}{{\∂w}_{\mathrm{ij}}}$

$=-\mathrm{\α}\frac{{\∂E}_k}{{\∂c}_{\mathrm{jnet}}}\frac{{\∂c}_{\mathrm{jnet}}}{{\∂w}_{\mathrm{ij}}}$

$={\mathrm{\αd}}_j\left[\frac{\∂\left(\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{b}_i\right)}{{\∂w}_{\mathrm{ij}}}\right]$

$={\mathrm{\αd}}_j{b}_i---(1-10)$

In like manner:

${\mathrm{\Δv}}_{\mathrm{hi}}=-\mathrm{\β}\frac{{\∂E}_k}{{\∂v}_{\mathrm{hi}}}$

$=-\mathrm{\β}\frac{{\∂E}_k}{{\∂b}_{\mathrm{inet}}}\frac{{\∂b}_{\mathrm{inet}}}{{\∂v}_{\mathrm{hi}}}$

$={\mathrm{\βe}}_i{\mathrm{\α}}_h---(1-11)$

In formula (1-10), (1-11), _α(0 < _α< 1) and β (0 < β < 1) be learning rate.

Because global cost function E is defined on whole training set, realize Gradient Descent real on E curved surface, need concentrate each pattern to during offering network at whole pattern drill, keep connection weight constant, obtain the negative gradient of E to connection weight, that is:

$-\frac{\&PartialD;E}{{\&PartialD;w}_{\mathrm{ij}}}=\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}(-\frac{{\&PartialD;E}_k}{{\&PartialD;w}_{\mathrm{ij}}})---(1-12)$

With:

$-\frac{\&PartialD;E}{{\&PartialD;v}_{\mathrm{hi}}}=\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}(-\frac{{\&PartialD;E}_k}{{\&PartialD;v}_{\mathrm{hi}}})---(1-13)$

Now, being changed to of connection weight:

${\mathrm{\&Delta;w}}_{\mathrm{ij}}=-\mathrm{\&alpha;}\frac{\&PartialD;E}{{\&PartialD;w}_{\mathrm{ij}}}$

$=\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}(-\mathrm{\&alpha;}\frac{{\&PartialD;E}_k}{{\&PartialD;w}_{\mathrm{ij}}})---(1-14)$

With:

${\mathrm{\&Delta;v}}_{\mathrm{hi}}=-\mathrm{\&beta;}\frac{\&PartialD;E}{{\&PartialD;w}_{\mathrm{ij}}}$

$=\underset{k=1}{\overset{p}{\mathrm{\&Sigma;}}}(-\mathrm{\&beta;}\frac{{\&PartialD;E}_k}{{\&PartialD;v}_{\mathrm{hi}}})---(1-15)$

From formula (1-14) and formula (1-15), can find out, connection weight variation is proportional to negative gradient sum corresponding to each pattern on whole set of patterns, and therefore the Gradient Descent of this algorithm has departed from the upper real Gradient Descent of E.But, enough hour of the learning rate α in formula (1-10) and formula (1-11) and β, this departing from is negligible.This pattern that often provides is commonly referred to the contrary propagation of standard error to just carrying out the learning algorithm of a connection weight adjustment.Above-described Back Propagation Algorithm is exactly the error Back-Propagation learning algorithm of standard.

Improved BP algorithm: because BP algorithm is based upon on functional gradient basis, restrain by the direction of error function Gradient Descent, this just unavoidably faces subject matter below:

1) excessive as converging factor β, can cause that vibration even disperses or converge on local minimum, when too hour, training can be very slow.

2) when training process approaches along with the minimal point to objective function, speed of convergence declines gradually, causes speed of convergence slow.

Improving BP algorithm is that gradient method and optimized thought are organically combined, and makes full use of t-1 effective information constantly, a kind of new optimizing algorithm of proposition.Its basic ideas are: if subsequent iteration makes for twice

symbol is identical, shows " decline a bit slow ", and step-length is too little, at this moment tackles its prize, increases step-length; Otherwise, as subsequent iteration makes for twice

symbol contrary, show " decline excessive ", step-length is too large, at this moment tackles it and penalizes, and reduces step-length.In order to make the unlikely amplitude of each change of step-length too large, this research adopts following linearity to encourage learning algorithm again:

Δη(t)=ελη(t-1) (1-16)

Wherein 0≤ε≤1 is constant, general desirable ε=0.20 ~ 0.30, and lambda definition is:

$\mathrm{\&lambda;}=\mathrm{sgn}(\frac{\&PartialD;E}{\&PartialD;\mathrm{BP}\left(t\right)}\&CenterDot;\frac{\&PartialD;E}{\&PartialD;\mathrm{BP}(t-1)})---(1-17)$

At this moment BP algorithm is:

$\mathrm{BP}(t+1)=\mathrm{BP}\left(t\right)-\mathrm{\&eta;}\left(t\right)\frac{\&PartialD;E}{\&PartialD;\mathrm{BP}\left(t\right)}---(1-18)$

Obtain, with the Adaptive Adjustment of Step Length Universal-purpose quick BP algorithm of momentum term, be:

BP(t+1)=BP(t)-η(t)Z(t) (1-19)

$Z\left(t\right)=\frac{\&PartialD;E}{\&PartialD;\mathrm{BP}\left(t\right)}+\mathrm{\&alpha;Z}(t-1)---(1-20)$

The form of being write the connection weights of each layer of neuroid and threshold value as above formula obtains the Adaptive Adjustment of Step Length rapid bp algorithm of three layers of error back propagation band momentum term:

W _hi(t+1)=W _hi(t)-η(t)Z(t) (1-21)

$Z\left(t\right)=\frac{\&PartialD;E}{{\&PartialD;W}_{\mathrm{hi}}\left(t\right)}+\mathrm{\&alpha;Z}(t-1)---(1-22)$

θ _i(t+1)=θ _i(t)-η(t)Z(t) (1-23)

$Z\left(t\right)=\frac{\&PartialD;E}{{\&PartialD;\mathrm{\&theta;}}_i\left(t\right)}+\mathrm{\&alpha;Z}(t-1)---(1-24)$

V _ij(t+1)＝V _ij(t)-η(t)Z(t) (1-25)

$Z\left(t\right)=\frac{\&PartialD;E}{{\&PartialD;V}_{\mathrm{ij}}\left(t\right)}+\mathrm{\&alpha;Z}(t-1)---(1-26)$

γ _j(t+1)＝γ _j(t)-η(t)Z(t) (1-27)

$Z\left(t\right)=\frac{\&PartialD;E}{{\&PartialD;\mathrm{\&gamma;}}_j\left(t\right)}+\mathrm{\&alpha;Z}(t-1)---(1-28)$

In formula: 0< α <1 is factor of momentum, 0< β <1 is the study factor, and t is the training time, and E is network energy function, W (t) is weight function, the function of Z (t) for introducing.

The self-adaptation variable step rapid bp algorithm that use is encouraged with the linearity of momentum term again, in learning process, make the adjusting of weights towards the mean direction variation of bottom, the swing that unlikely generation is large, play the effect of buffering and smoothing, if system enters the flat region of function surface, error will change very littlely so, so Δ W (t+1) is similar to Δ W (t), and average Δ W will become:

in formula what become is more effective, makes to regulate to depart from as early as possible saturation region.It is low and easily converge on the problem of local minimum that this can effectively solve learning efficiency.

Fig. 3 has disclosed according to the model learning flow process based on improving BP algorithm in the BF Slag Viscosity forecasting procedure of a specific implementation of the present invention.Viscosity forecasting model is the model based on artificial neural network, adopts improved BP algorithm, is divided into foundation and the viscosity forecast of Performance of Slag database and analyzes two large modules.Model has the functions such as data query, data derivation, slag composition trend analysis, Model Self-Learning training, Performance of Slag forecast, forecast result analysis, help system.The learning process of model is exactly to connecting the adjustment process of weights and threshold value.Connection weight usually can not be determined in advance exactly, also can adjust gradually weights according to sample mode, makes neuroid have the function of remarkable process information.What this model adopted is improved BP algorithm, has pace of learning faster.

The setting of study situation: in BP model topological structure, the neuron number of input and output is determined by problem itself.For the number of hidden layer, and hidden neuron keeps count of definite be the problem of a more complicated, not yet have at present a governing principle.This model use neuroid is determined Hidden unit number from configuration algorithm.This algorithm utilizes the method for mathematical statistics to introduce correlativity and the sample dispersion degree concept between Hidden unit, consider the contribution of hidden neuron to network, delete as calculated or to merge those effects little or act on close neuron, thereby determine rational network structure.Concrete grammar is as follows:

If O _ipthe output of hidden neuron i when p sample of study,

be the average output of hidden first i after having learnt n sample, n is the total sample number of training:

sample dispersion degree

if | S _i| <C ₁, C ₁get 0.001～0.01, illustrate that the exporting change of hidden first i is very little, can delete.Whether two can merge with the hidden node i of layer and j, will weigh their related coefficient:

$r_{\mathrm{ij}}=\frac{\frac{1}{n}\underset{p=1}{\overset{n}{\mathrm{\&Sigma;}}}{O}_{\mathrm{ip}}{O}_{\mathrm{jp}}-{\stackrel{\&OverBar;}{O}}_i{\stackrel{\&OverBar;}{O}}_j}{S_i{S}_j}---(1-29)$

If r _ij>=C ₂, C ₂get 0.8～0.9, illustrate that the function of i and j repeats, can unite two into one.According to this method, select for the maximally related mode input parameter of slag viscosity, and tentatively determine hidden layer neuron number, then pass through repeatedly forecast experiments and adjust parameter, finally determine rational network model parameter.

The study of model: according to the performance parameter of blast furnace slag (as SiO ₂, Al ₂o ₃, CaO, MgO percentage composition, η _{1500 ℃}, η _{1475 ℃}, η _{1450 ℃}, Tm) condition, the learning data in garbled data storehouse, sets learning parameter, then just can carry out the study of model.After model learning finishes, can automatically preserve the result after study, connect weights and threshold value, then just can carry out Performance of Slag forecast.

Performance of Slag forecasting model and the contrast of traditional experimental formula error:

All there is certain error in Performance of Slag model prediction and experimental formula, below both errors is contrasted.The experimental formula that this contrast adopts is by 45 groups of experimental datas are carried out to the multiple linear formula that multiple linear regression obtains, as follows:

η _1500℃=0.105+0.012w(Al ₂O ₃)％-0.0034w(MgO)％-0.008R ₄

R=0.912

η _1475℃=0.17+0.017w(Al ₂O ₃)％-0.005w(MgO)％-0.11R ₄

R=0.931

η _1450℃=0.22+0.024w(Al ₂O ₃)％-0.008w(MgO)％-0.18R ₄

R=0.914

Tm=1246.5+10.232w(Al ₂O ₃)％-3.779w(MgO)％-11.546R ₄

R=0.972

The formula scope of application: R ₂for (1.1～1.3), Al ₂o ₃content is that (15%～19%), MgO content are (7%～11%).

Fig. 4 and Fig. 5 have disclosed respectively the error statistics that adopts BF Slag Viscosity forecasting procedure of the present invention and adopt traditional linear regression formula mode to forecast.Wherein Fig. 4 has disclosed the error statistics that adopts BF Slag Viscosity forecasting procedure of the present invention.Fig. 5 has disclosed the error statistics that adopts traditional linear regression formula mode to forecast.By error, contrast and can find out, the error of Performance of Slag model prediction is basicly stable between ± 2%, the error of linear regression formula is stabilized in ± 6% between, the Performance of Slag forecasting model of explanation based on neuroid has higher accuracy than traditional experimental formula method, neural network model also has self-learning capability simultaneously, and the composition range of forecast is wider.

According to practical application test of the present invention, obtain the result contrast of actual viscosity measured value and forecast, as shown in table 1 below:

Table 1 slag viscosity measured value and predicted value contrast

By error information mapping analysis, as follows

From upper table and upper figure, can find out, for on-the-spot slag, the error of system forecast is substantially all in 10%, and hit rate, more than 90%, can meet the accuracy requirement of blast furnace process to slag parameter.

Simultaneity factor forecast has obtained good effect in blast furnace is produced, and blast furnace indices, comprises that operation index and Performance of Slag parameter all increase, as shown in following table 2 and table 3:

Table 2 operation index

Table 3 Performance of Slag parameter

By above table 2 and table 3, can be found out:

(1) lump ore ratio increases

Rationally use lump ore can reduce ironmaking cost, improve the market competitiveness of enterprise.Use after native system, lump ore ratio brings up to 20% by 12%, and pellet is reduced to 5% by 17%, and cost is very considerable.

(2) MgO content reduction, the quantity of slag reduce

Mostly lump ore is Al ₂o ₃the ore deposit that content is higher, furnace charge lump ore ratio increases, and can cause Al in slag ₂o ₃content raises, and viscosity can increase.In order to reduce slag viscosity, blast furnace generally adopts the method that increases MgO content both at home and abroad now, can cause so again the quantity of slag of slag to increase.Before not using the present invention, MgO content is not more than 9%, and about slag ratio 300Kg/t, the present invention is by the reasonably optimizing that slag is formed, make MgO content be reduced to 8% left and right, slag ratio drops to 275Kg/t left and right, and the ratio of first-grade products of molten iron still remains on 98% left and right, although Al ₂o ₃content increases to 16～17% left and right by 14%, and slag still can keep good mobility.

(3) coke ratio reduction, coal are than improving

Because slag ratio is larger, blast furnace coke ratio is 410Kg/t left and right always, and coal ratio is about 100Kg/t left and right.The present invention can optimize slag system and form, and slag ratio reduces, and the coke ratio of blast furnace is reduced to 370Kg/t left and right, and coal ratio is brought up to 150Kg/t left and right.

Adopt a concrete learning process of pre-guarantor's method of the present invention as follows:

(1) initial setting up connection weight (w _j, v _i), threshold value (θ _i) and transport function f ();

(2) input sample data, calculates hidden layer output (b _i) and output layer output (c _j), $b_i=f\left(b_{\mathrm{inet}}\right)=f\left(\underset{h=1}{\overset{L}{\mathrm{\&Sigma;}}}{v}_{\mathrm{hi}}{a}_h\right),$ $c_j=f\left(c_{\mathrm{jnet}}\right)=f\left(\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{b}_i\right)$

(3) calculate output layer error (d _j) and hidden layer error (e _i),

$e_i=f^{\&prime;}\left(b_{\mathrm{inet}}\right)\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{d}_j{w}_{\mathrm{ij}};$

(4) calculate each layer of gradient, set α, β, and revise weights; Δ w _ij=α d _jb _i, Δ v _hi=β e _iα _h;

(5) connection weight (w _j, v _i), threshold value (θ _i) increase step-length,

W _hi(t+1)=W _hi(t)η(t)Z(t)、

θ _i(t+1)=θ _i(t)-η(t)Z(t)、

V _ij(t+1)=V _ij(t)-η(t)Z(t)、

(6) by the connection weight (w increasing after step-length _j(t+1), v _i) and threshold value (θ (t+1) _i(t+1)) replace original connection weight (w _j, v _i) and threshold value (θ _i)

(7) input sample data, from step (2), repeats said process.

(8) when error, be less than the accuracy value of setting, stop calculating, preserve final connection weight (w _j, v _i) and threshold value (θ _i).

By the result of study, be connection weight (w _j, v _i) and threshold value (θ _i), after preservation, represent that study finishes.

Although above, native system is described in detail, to those skilled in the art, can expand the input and output of system.For example input increases some other slag compositions as TiO ₂, FeO, MnO content etc.; Output increases take expertise as basic composition range early warning and improves countermeasure etc.Native system can also be combined with sintered material system, compared with the prediction blast furnace process situation of system, for batching provides theoretical foundation more accurately.The neural network model method that native system adopts can also be with in other respects, as sintering deposit, pellet composition and metallurgical performance forecast, molten iron Si content (molten iron temperature) forecast, steelmaking slag, refining slag performance prediction etc.

The BF Slag Viscosity forecasting procedure that the present invention proposes realizes based on neuroid, can be stabilized in ± 2% left and right of error, than traditional error ± 6% based on linear regression formula mode, be significantly improved, neuroid has self-learning capability simultaneously, can use the scope of forecasting widely.

Claims (4)

1. a BF Slag Viscosity forecasting procedure, is characterized in that, comprising:

Neuroid construction step, the neuroid structure building comprises input layer, hidden layer and output layer, between input layer, hidden layer and output layer, by connecting weights, connect, hidden layer has threshold value, the input parameter of input layer is the percentage composition of each composition in slag, and the output parameter of output layer is slag viscosity, melting temperature and the desulfurization coefficient under different temperatures;

Initialization step, obtains training data, and to training data, connect weights and threshold value is carried out initialization, described training data comprises reference input parameter and with reference to output parameter;

Neuroid self study step, extract a training data, reference input parameter is wherein imported to neuroid, via input layer, hidden layer and output layer output training output parameter, relatively train the reference output parameter in output parameter and this training data, calculate error and the partial gradient of each layer, according to error, to set step-length adjustment, be connected weights and threshold value with partial gradient, according to partial gradient adjustment, set step-length, extract next training data and repeat said process, until all training datas are all used, preserve the end value that connects weights and threshold value,

Slag viscosity forecast step, is fixed on end value by connection weights and the threshold value of neuroid, inputs the percentage composition of each composition in actual slag, slag viscosity, melting temperature and desulfurization coefficient under the different temperatures of output forecast.

2. BF Slag Viscosity forecasting procedure as claimed in claim 1, is characterized in that,

In described slag, the percentage composition of each composition comprises SiO ₂, Al ₂the percentage composition of O3, CaO, MgO;

Slag viscosity under described different temperatures comprises 1450 ℃ of the slag viscosity η of 1475 ℃ of the slag viscosity η of 1500 ℃, 1475 ℃ of slag viscosity η of 1500 ℃ and 1450 ℃.

3. BF Slag Viscosity forecasting procedure as claimed in claim 1, it is characterized in that, training data is carried out to initialization and comprise training data is normalized, to connecting weights, carry out initialization with threshold value and comprise that by being connected weights and Threshold be initial value.

4. BF Slag Viscosity forecasting procedure as claimed in claim 1, is characterized in that, adjusts described setting step-length comprise according to partial gradient:

To partial gradient subsequent iteration twice, if two sub-symbols are identical, increase and set step-length; If two sub-symbols are contrary, reduce to set step-length.

Images